~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.16-rc3 ] ~ [ linux-5.15.5 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.82 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.162 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.218 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.256 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.291 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.293 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0-only
  2 /*
  3  * Generic hugetlb support.
  4  * (C) Nadia Yvette Chambers, April 2004
  5  */
  6 #include <linux/list.h>
  7 #include <linux/init.h>
  8 #include <linux/mm.h>
  9 #include <linux/seq_file.h>
 10 #include <linux/sysctl.h>
 11 #include <linux/highmem.h>
 12 #include <linux/mmu_notifier.h>
 13 #include <linux/nodemask.h>
 14 #include <linux/pagemap.h>
 15 #include <linux/mempolicy.h>
 16 #include <linux/compiler.h>
 17 #include <linux/cpuset.h>
 18 #include <linux/mutex.h>
 19 #include <linux/memblock.h>
 20 #include <linux/sysfs.h>
 21 #include <linux/slab.h>
 22 #include <linux/mmdebug.h>
 23 #include <linux/sched/signal.h>
 24 #include <linux/rmap.h>
 25 #include <linux/string_helpers.h>
 26 #include <linux/swap.h>
 27 #include <linux/swapops.h>
 28 #include <linux/jhash.h>
 29 #include <linux/numa.h>
 30 
 31 #include <asm/page.h>
 32 #include <asm/pgtable.h>
 33 #include <asm/tlb.h>
 34 
 35 #include <linux/io.h>
 36 #include <linux/hugetlb.h>
 37 #include <linux/hugetlb_cgroup.h>
 38 #include <linux/node.h>
 39 #include <linux/userfaultfd_k.h>
 40 #include <linux/page_owner.h>
 41 #include "internal.h"
 42 
 43 int hugetlb_max_hstate __read_mostly;
 44 unsigned int default_hstate_idx;
 45 struct hstate hstates[HUGE_MAX_HSTATE];
 46 /*
 47  * Minimum page order among possible hugepage sizes, set to a proper value
 48  * at boot time.
 49  */
 50 static unsigned int minimum_order __read_mostly = UINT_MAX;
 51 
 52 __initdata LIST_HEAD(huge_boot_pages);
 53 
 54 /* for command line parsing */
 55 static struct hstate * __initdata parsed_hstate;
 56 static unsigned long __initdata default_hstate_max_huge_pages;
 57 static unsigned long __initdata default_hstate_size;
 58 static bool __initdata parsed_valid_hugepagesz = true;
 59 
 60 /*
 61  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
 62  * free_huge_pages, and surplus_huge_pages.
 63  */
 64 DEFINE_SPINLOCK(hugetlb_lock);
 65 
 66 /*
 67  * Serializes faults on the same logical page.  This is used to
 68  * prevent spurious OOMs when the hugepage pool is fully utilized.
 69  */
 70 static int num_fault_mutexes;
 71 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
 72 
 73 /* Forward declaration */
 74 static int hugetlb_acct_memory(struct hstate *h, long delta);
 75 
 76 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
 77 {
 78         bool free = (spool->count == 0) && (spool->used_hpages == 0);
 79 
 80         spin_unlock(&spool->lock);
 81 
 82         /* If no pages are used, and no other handles to the subpool
 83          * remain, give up any reservations mased on minimum size and
 84          * free the subpool */
 85         if (free) {
 86                 if (spool->min_hpages != -1)
 87                         hugetlb_acct_memory(spool->hstate,
 88                                                 -spool->min_hpages);
 89                 kfree(spool);
 90         }
 91 }
 92 
 93 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
 94                                                 long min_hpages)
 95 {
 96         struct hugepage_subpool *spool;
 97 
 98         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
 99         if (!spool)
100                 return NULL;
101 
102         spin_lock_init(&spool->lock);
103         spool->count = 1;
104         spool->max_hpages = max_hpages;
105         spool->hstate = h;
106         spool->min_hpages = min_hpages;
107 
108         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
109                 kfree(spool);
110                 return NULL;
111         }
112         spool->rsv_hpages = min_hpages;
113 
114         return spool;
115 }
116 
117 void hugepage_put_subpool(struct hugepage_subpool *spool)
118 {
119         spin_lock(&spool->lock);
120         BUG_ON(!spool->count);
121         spool->count--;
122         unlock_or_release_subpool(spool);
123 }
124 
125 /*
126  * Subpool accounting for allocating and reserving pages.
127  * Return -ENOMEM if there are not enough resources to satisfy the
128  * the request.  Otherwise, return the number of pages by which the
129  * global pools must be adjusted (upward).  The returned value may
130  * only be different than the passed value (delta) in the case where
131  * a subpool minimum size must be manitained.
132  */
133 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
134                                       long delta)
135 {
136         long ret = delta;
137 
138         if (!spool)
139                 return ret;
140 
141         spin_lock(&spool->lock);
142 
143         if (spool->max_hpages != -1) {          /* maximum size accounting */
144                 if ((spool->used_hpages + delta) <= spool->max_hpages)
145                         spool->used_hpages += delta;
146                 else {
147                         ret = -ENOMEM;
148                         goto unlock_ret;
149                 }
150         }
151 
152         /* minimum size accounting */
153         if (spool->min_hpages != -1 && spool->rsv_hpages) {
154                 if (delta > spool->rsv_hpages) {
155                         /*
156                          * Asking for more reserves than those already taken on
157                          * behalf of subpool.  Return difference.
158                          */
159                         ret = delta - spool->rsv_hpages;
160                         spool->rsv_hpages = 0;
161                 } else {
162                         ret = 0;        /* reserves already accounted for */
163                         spool->rsv_hpages -= delta;
164                 }
165         }
166 
167 unlock_ret:
168         spin_unlock(&spool->lock);
169         return ret;
170 }
171 
172 /*
173  * Subpool accounting for freeing and unreserving pages.
174  * Return the number of global page reservations that must be dropped.
175  * The return value may only be different than the passed value (delta)
176  * in the case where a subpool minimum size must be maintained.
177  */
178 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
179                                        long delta)
180 {
181         long ret = delta;
182 
183         if (!spool)
184                 return delta;
185 
186         spin_lock(&spool->lock);
187 
188         if (spool->max_hpages != -1)            /* maximum size accounting */
189                 spool->used_hpages -= delta;
190 
191          /* minimum size accounting */
192         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
193                 if (spool->rsv_hpages + delta <= spool->min_hpages)
194                         ret = 0;
195                 else
196                         ret = spool->rsv_hpages + delta - spool->min_hpages;
197 
198                 spool->rsv_hpages += delta;
199                 if (spool->rsv_hpages > spool->min_hpages)
200                         spool->rsv_hpages = spool->min_hpages;
201         }
202 
203         /*
204          * If hugetlbfs_put_super couldn't free spool due to an outstanding
205          * quota reference, free it now.
206          */
207         unlock_or_release_subpool(spool);
208 
209         return ret;
210 }
211 
212 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
213 {
214         return HUGETLBFS_SB(inode->i_sb)->spool;
215 }
216 
217 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
218 {
219         return subpool_inode(file_inode(vma->vm_file));
220 }
221 
222 /*
223  * Region tracking -- allows tracking of reservations and instantiated pages
224  *                    across the pages in a mapping.
225  *
226  * The region data structures are embedded into a resv_map and protected
227  * by a resv_map's lock.  The set of regions within the resv_map represent
228  * reservations for huge pages, or huge pages that have already been
229  * instantiated within the map.  The from and to elements are huge page
230  * indicies into the associated mapping.  from indicates the starting index
231  * of the region.  to represents the first index past the end of  the region.
232  *
233  * For example, a file region structure with from == 0 and to == 4 represents
234  * four huge pages in a mapping.  It is important to note that the to element
235  * represents the first element past the end of the region. This is used in
236  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
237  *
238  * Interval notation of the form [from, to) will be used to indicate that
239  * the endpoint from is inclusive and to is exclusive.
240  */
241 struct file_region {
242         struct list_head link;
243         long from;
244         long to;
245 };
246 
247 /*
248  * Add the huge page range represented by [f, t) to the reserve
249  * map.  In the normal case, existing regions will be expanded
250  * to accommodate the specified range.  Sufficient regions should
251  * exist for expansion due to the previous call to region_chg
252  * with the same range.  However, it is possible that region_del
253  * could have been called after region_chg and modifed the map
254  * in such a way that no region exists to be expanded.  In this
255  * case, pull a region descriptor from the cache associated with
256  * the map and use that for the new range.
257  *
258  * Return the number of new huge pages added to the map.  This
259  * number is greater than or equal to zero.
260  */
261 static long region_add(struct resv_map *resv, long f, long t)
262 {
263         struct list_head *head = &resv->regions;
264         struct file_region *rg, *nrg, *trg;
265         long add = 0;
266 
267         spin_lock(&resv->lock);
268         /* Locate the region we are either in or before. */
269         list_for_each_entry(rg, head, link)
270                 if (f <= rg->to)
271                         break;
272 
273         /*
274          * If no region exists which can be expanded to include the
275          * specified range, the list must have been modified by an
276          * interleving call to region_del().  Pull a region descriptor
277          * from the cache and use it for this range.
278          */
279         if (&rg->link == head || t < rg->from) {
280                 VM_BUG_ON(resv->region_cache_count <= 0);
281 
282                 resv->region_cache_count--;
283                 nrg = list_first_entry(&resv->region_cache, struct file_region,
284                                         link);
285                 list_del(&nrg->link);
286 
287                 nrg->from = f;
288                 nrg->to = t;
289                 list_add(&nrg->link, rg->link.prev);
290 
291                 add += t - f;
292                 goto out_locked;
293         }
294 
295         /* Round our left edge to the current segment if it encloses us. */
296         if (f > rg->from)
297                 f = rg->from;
298 
299         /* Check for and consume any regions we now overlap with. */
300         nrg = rg;
301         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
302                 if (&rg->link == head)
303                         break;
304                 if (rg->from > t)
305                         break;
306 
307                 /* If this area reaches higher then extend our area to
308                  * include it completely.  If this is not the first area
309                  * which we intend to reuse, free it. */
310                 if (rg->to > t)
311                         t = rg->to;
312                 if (rg != nrg) {
313                         /* Decrement return value by the deleted range.
314                          * Another range will span this area so that by
315                          * end of routine add will be >= zero
316                          */
317                         add -= (rg->to - rg->from);
318                         list_del(&rg->link);
319                         kfree(rg);
320                 }
321         }
322 
323         add += (nrg->from - f);         /* Added to beginning of region */
324         nrg->from = f;
325         add += t - nrg->to;             /* Added to end of region */
326         nrg->to = t;
327 
328 out_locked:
329         resv->adds_in_progress--;
330         spin_unlock(&resv->lock);
331         VM_BUG_ON(add < 0);
332         return add;
333 }
334 
335 /*
336  * Examine the existing reserve map and determine how many
337  * huge pages in the specified range [f, t) are NOT currently
338  * represented.  This routine is called before a subsequent
339  * call to region_add that will actually modify the reserve
340  * map to add the specified range [f, t).  region_chg does
341  * not change the number of huge pages represented by the
342  * map.  However, if the existing regions in the map can not
343  * be expanded to represent the new range, a new file_region
344  * structure is added to the map as a placeholder.  This is
345  * so that the subsequent region_add call will have all the
346  * regions it needs and will not fail.
347  *
348  * Upon entry, region_chg will also examine the cache of region descriptors
349  * associated with the map.  If there are not enough descriptors cached, one
350  * will be allocated for the in progress add operation.
351  *
352  * Returns the number of huge pages that need to be added to the existing
353  * reservation map for the range [f, t).  This number is greater or equal to
354  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
355  * is needed and can not be allocated.
356  */
357 static long region_chg(struct resv_map *resv, long f, long t)
358 {
359         struct list_head *head = &resv->regions;
360         struct file_region *rg, *nrg = NULL;
361         long chg = 0;
362 
363 retry:
364         spin_lock(&resv->lock);
365 retry_locked:
366         resv->adds_in_progress++;
367 
368         /*
369          * Check for sufficient descriptors in the cache to accommodate
370          * the number of in progress add operations.
371          */
372         if (resv->adds_in_progress > resv->region_cache_count) {
373                 struct file_region *trg;
374 
375                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
376                 /* Must drop lock to allocate a new descriptor. */
377                 resv->adds_in_progress--;
378                 spin_unlock(&resv->lock);
379 
380                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
381                 if (!trg) {
382                         kfree(nrg);
383                         return -ENOMEM;
384                 }
385 
386                 spin_lock(&resv->lock);
387                 list_add(&trg->link, &resv->region_cache);
388                 resv->region_cache_count++;
389                 goto retry_locked;
390         }
391 
392         /* Locate the region we are before or in. */
393         list_for_each_entry(rg, head, link)
394                 if (f <= rg->to)
395                         break;
396 
397         /* If we are below the current region then a new region is required.
398          * Subtle, allocate a new region at the position but make it zero
399          * size such that we can guarantee to record the reservation. */
400         if (&rg->link == head || t < rg->from) {
401                 if (!nrg) {
402                         resv->adds_in_progress--;
403                         spin_unlock(&resv->lock);
404                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
405                         if (!nrg)
406                                 return -ENOMEM;
407 
408                         nrg->from = f;
409                         nrg->to   = f;
410                         INIT_LIST_HEAD(&nrg->link);
411                         goto retry;
412                 }
413 
414                 list_add(&nrg->link, rg->link.prev);
415                 chg = t - f;
416                 goto out_nrg;
417         }
418 
419         /* Round our left edge to the current segment if it encloses us. */
420         if (f > rg->from)
421                 f = rg->from;
422         chg = t - f;
423 
424         /* Check for and consume any regions we now overlap with. */
425         list_for_each_entry(rg, rg->link.prev, link) {
426                 if (&rg->link == head)
427                         break;
428                 if (rg->from > t)
429                         goto out;
430 
431                 /* We overlap with this area, if it extends further than
432                  * us then we must extend ourselves.  Account for its
433                  * existing reservation. */
434                 if (rg->to > t) {
435                         chg += rg->to - t;
436                         t = rg->to;
437                 }
438                 chg -= rg->to - rg->from;
439         }
440 
441 out:
442         spin_unlock(&resv->lock);
443         /*  We already know we raced and no longer need the new region */
444         kfree(nrg);
445         return chg;
446 out_nrg:
447         spin_unlock(&resv->lock);
448         return chg;
449 }
450 
451 /*
452  * Abort the in progress add operation.  The adds_in_progress field
453  * of the resv_map keeps track of the operations in progress between
454  * calls to region_chg and region_add.  Operations are sometimes
455  * aborted after the call to region_chg.  In such cases, region_abort
456  * is called to decrement the adds_in_progress counter.
457  *
458  * NOTE: The range arguments [f, t) are not needed or used in this
459  * routine.  They are kept to make reading the calling code easier as
460  * arguments will match the associated region_chg call.
461  */
462 static void region_abort(struct resv_map *resv, long f, long t)
463 {
464         spin_lock(&resv->lock);
465         VM_BUG_ON(!resv->region_cache_count);
466         resv->adds_in_progress--;
467         spin_unlock(&resv->lock);
468 }
469 
470 /*
471  * Delete the specified range [f, t) from the reserve map.  If the
472  * t parameter is LONG_MAX, this indicates that ALL regions after f
473  * should be deleted.  Locate the regions which intersect [f, t)
474  * and either trim, delete or split the existing regions.
475  *
476  * Returns the number of huge pages deleted from the reserve map.
477  * In the normal case, the return value is zero or more.  In the
478  * case where a region must be split, a new region descriptor must
479  * be allocated.  If the allocation fails, -ENOMEM will be returned.
480  * NOTE: If the parameter t == LONG_MAX, then we will never split
481  * a region and possibly return -ENOMEM.  Callers specifying
482  * t == LONG_MAX do not need to check for -ENOMEM error.
483  */
484 static long region_del(struct resv_map *resv, long f, long t)
485 {
486         struct list_head *head = &resv->regions;
487         struct file_region *rg, *trg;
488         struct file_region *nrg = NULL;
489         long del = 0;
490 
491 retry:
492         spin_lock(&resv->lock);
493         list_for_each_entry_safe(rg, trg, head, link) {
494                 /*
495                  * Skip regions before the range to be deleted.  file_region
496                  * ranges are normally of the form [from, to).  However, there
497                  * may be a "placeholder" entry in the map which is of the form
498                  * (from, to) with from == to.  Check for placeholder entries
499                  * at the beginning of the range to be deleted.
500                  */
501                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
502                         continue;
503 
504                 if (rg->from >= t)
505                         break;
506 
507                 if (f > rg->from && t < rg->to) { /* Must split region */
508                         /*
509                          * Check for an entry in the cache before dropping
510                          * lock and attempting allocation.
511                          */
512                         if (!nrg &&
513                             resv->region_cache_count > resv->adds_in_progress) {
514                                 nrg = list_first_entry(&resv->region_cache,
515                                                         struct file_region,
516                                                         link);
517                                 list_del(&nrg->link);
518                                 resv->region_cache_count--;
519                         }
520 
521                         if (!nrg) {
522                                 spin_unlock(&resv->lock);
523                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
524                                 if (!nrg)
525                                         return -ENOMEM;
526                                 goto retry;
527                         }
528 
529                         del += t - f;
530 
531                         /* New entry for end of split region */
532                         nrg->from = t;
533                         nrg->to = rg->to;
534                         INIT_LIST_HEAD(&nrg->link);
535 
536                         /* Original entry is trimmed */
537                         rg->to = f;
538 
539                         list_add(&nrg->link, &rg->link);
540                         nrg = NULL;
541                         break;
542                 }
543 
544                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
545                         del += rg->to - rg->from;
546                         list_del(&rg->link);
547                         kfree(rg);
548                         continue;
549                 }
550 
551                 if (f <= rg->from) {    /* Trim beginning of region */
552                         del += t - rg->from;
553                         rg->from = t;
554                 } else {                /* Trim end of region */
555                         del += rg->to - f;
556                         rg->to = f;
557                 }
558         }
559 
560         spin_unlock(&resv->lock);
561         kfree(nrg);
562         return del;
563 }
564 
565 /*
566  * A rare out of memory error was encountered which prevented removal of
567  * the reserve map region for a page.  The huge page itself was free'ed
568  * and removed from the page cache.  This routine will adjust the subpool
569  * usage count, and the global reserve count if needed.  By incrementing
570  * these counts, the reserve map entry which could not be deleted will
571  * appear as a "reserved" entry instead of simply dangling with incorrect
572  * counts.
573  */
574 void hugetlb_fix_reserve_counts(struct inode *inode)
575 {
576         struct hugepage_subpool *spool = subpool_inode(inode);
577         long rsv_adjust;
578 
579         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
580         if (rsv_adjust) {
581                 struct hstate *h = hstate_inode(inode);
582 
583                 hugetlb_acct_memory(h, 1);
584         }
585 }
586 
587 /*
588  * Count and return the number of huge pages in the reserve map
589  * that intersect with the range [f, t).
590  */
591 static long region_count(struct resv_map *resv, long f, long t)
592 {
593         struct list_head *head = &resv->regions;
594         struct file_region *rg;
595         long chg = 0;
596 
597         spin_lock(&resv->lock);
598         /* Locate each segment we overlap with, and count that overlap. */
599         list_for_each_entry(rg, head, link) {
600                 long seg_from;
601                 long seg_to;
602 
603                 if (rg->to <= f)
604                         continue;
605                 if (rg->from >= t)
606                         break;
607 
608                 seg_from = max(rg->from, f);
609                 seg_to = min(rg->to, t);
610 
611                 chg += seg_to - seg_from;
612         }
613         spin_unlock(&resv->lock);
614 
615         return chg;
616 }
617 
618 /*
619  * Convert the address within this vma to the page offset within
620  * the mapping, in pagecache page units; huge pages here.
621  */
622 static pgoff_t vma_hugecache_offset(struct hstate *h,
623                         struct vm_area_struct *vma, unsigned long address)
624 {
625         return ((address - vma->vm_start) >> huge_page_shift(h)) +
626                         (vma->vm_pgoff >> huge_page_order(h));
627 }
628 
629 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
630                                      unsigned long address)
631 {
632         return vma_hugecache_offset(hstate_vma(vma), vma, address);
633 }
634 EXPORT_SYMBOL_GPL(linear_hugepage_index);
635 
636 /*
637  * Return the size of the pages allocated when backing a VMA. In the majority
638  * cases this will be same size as used by the page table entries.
639  */
640 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
641 {
642         if (vma->vm_ops && vma->vm_ops->pagesize)
643                 return vma->vm_ops->pagesize(vma);
644         return PAGE_SIZE;
645 }
646 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
647 
648 /*
649  * Return the page size being used by the MMU to back a VMA. In the majority
650  * of cases, the page size used by the kernel matches the MMU size. On
651  * architectures where it differs, an architecture-specific 'strong'
652  * version of this symbol is required.
653  */
654 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
655 {
656         return vma_kernel_pagesize(vma);
657 }
658 
659 /*
660  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
661  * bits of the reservation map pointer, which are always clear due to
662  * alignment.
663  */
664 #define HPAGE_RESV_OWNER    (1UL << 0)
665 #define HPAGE_RESV_UNMAPPED (1UL << 1)
666 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
667 
668 /*
669  * These helpers are used to track how many pages are reserved for
670  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671  * is guaranteed to have their future faults succeed.
672  *
673  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674  * the reserve counters are updated with the hugetlb_lock held. It is safe
675  * to reset the VMA at fork() time as it is not in use yet and there is no
676  * chance of the global counters getting corrupted as a result of the values.
677  *
678  * The private mapping reservation is represented in a subtly different
679  * manner to a shared mapping.  A shared mapping has a region map associated
680  * with the underlying file, this region map represents the backing file
681  * pages which have ever had a reservation assigned which this persists even
682  * after the page is instantiated.  A private mapping has a region map
683  * associated with the original mmap which is attached to all VMAs which
684  * reference it, this region map represents those offsets which have consumed
685  * reservation ie. where pages have been instantiated.
686  */
687 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688 {
689         return (unsigned long)vma->vm_private_data;
690 }
691 
692 static void set_vma_private_data(struct vm_area_struct *vma,
693                                                         unsigned long value)
694 {
695         vma->vm_private_data = (void *)value;
696 }
697 
698 struct resv_map *resv_map_alloc(void)
699 {
700         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
701         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702 
703         if (!resv_map || !rg) {
704                 kfree(resv_map);
705                 kfree(rg);
706                 return NULL;
707         }
708 
709         kref_init(&resv_map->refs);
710         spin_lock_init(&resv_map->lock);
711         INIT_LIST_HEAD(&resv_map->regions);
712 
713         resv_map->adds_in_progress = 0;
714 
715         INIT_LIST_HEAD(&resv_map->region_cache);
716         list_add(&rg->link, &resv_map->region_cache);
717         resv_map->region_cache_count = 1;
718 
719         return resv_map;
720 }
721 
722 void resv_map_release(struct kref *ref)
723 {
724         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
725         struct list_head *head = &resv_map->region_cache;
726         struct file_region *rg, *trg;
727 
728         /* Clear out any active regions before we release the map. */
729         region_del(resv_map, 0, LONG_MAX);
730 
731         /* ... and any entries left in the cache */
732         list_for_each_entry_safe(rg, trg, head, link) {
733                 list_del(&rg->link);
734                 kfree(rg);
735         }
736 
737         VM_BUG_ON(resv_map->adds_in_progress);
738 
739         kfree(resv_map);
740 }
741 
742 static inline struct resv_map *inode_resv_map(struct inode *inode)
743 {
744         /*
745          * At inode evict time, i_mapping may not point to the original
746          * address space within the inode.  This original address space
747          * contains the pointer to the resv_map.  So, always use the
748          * address space embedded within the inode.
749          * The VERY common case is inode->mapping == &inode->i_data but,
750          * this may not be true for device special inodes.
751          */
752         return (struct resv_map *)(&inode->i_data)->private_data;
753 }
754 
755 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
756 {
757         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
758         if (vma->vm_flags & VM_MAYSHARE) {
759                 struct address_space *mapping = vma->vm_file->f_mapping;
760                 struct inode *inode = mapping->host;
761 
762                 return inode_resv_map(inode);
763 
764         } else {
765                 return (struct resv_map *)(get_vma_private_data(vma) &
766                                                         ~HPAGE_RESV_MASK);
767         }
768 }
769 
770 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
771 {
772         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
774 
775         set_vma_private_data(vma, (get_vma_private_data(vma) &
776                                 HPAGE_RESV_MASK) | (unsigned long)map);
777 }
778 
779 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
780 {
781         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
783 
784         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
785 }
786 
787 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
788 {
789         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
790 
791         return (get_vma_private_data(vma) & flag) != 0;
792 }
793 
794 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
795 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
796 {
797         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
798         if (!(vma->vm_flags & VM_MAYSHARE))
799                 vma->vm_private_data = (void *)0;
800 }
801 
802 /* Returns true if the VMA has associated reserve pages */
803 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
804 {
805         if (vma->vm_flags & VM_NORESERVE) {
806                 /*
807                  * This address is already reserved by other process(chg == 0),
808                  * so, we should decrement reserved count. Without decrementing,
809                  * reserve count remains after releasing inode, because this
810                  * allocated page will go into page cache and is regarded as
811                  * coming from reserved pool in releasing step.  Currently, we
812                  * don't have any other solution to deal with this situation
813                  * properly, so add work-around here.
814                  */
815                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
816                         return true;
817                 else
818                         return false;
819         }
820 
821         /* Shared mappings always use reserves */
822         if (vma->vm_flags & VM_MAYSHARE) {
823                 /*
824                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
825                  * be a region map for all pages.  The only situation where
826                  * there is no region map is if a hole was punched via
827                  * fallocate.  In this case, there really are no reverves to
828                  * use.  This situation is indicated if chg != 0.
829                  */
830                 if (chg)
831                         return false;
832                 else
833                         return true;
834         }
835 
836         /*
837          * Only the process that called mmap() has reserves for
838          * private mappings.
839          */
840         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
841                 /*
842                  * Like the shared case above, a hole punch or truncate
843                  * could have been performed on the private mapping.
844                  * Examine the value of chg to determine if reserves
845                  * actually exist or were previously consumed.
846                  * Very Subtle - The value of chg comes from a previous
847                  * call to vma_needs_reserves().  The reserve map for
848                  * private mappings has different (opposite) semantics
849                  * than that of shared mappings.  vma_needs_reserves()
850                  * has already taken this difference in semantics into
851                  * account.  Therefore, the meaning of chg is the same
852                  * as in the shared case above.  Code could easily be
853                  * combined, but keeping it separate draws attention to
854                  * subtle differences.
855                  */
856                 if (chg)
857                         return false;
858                 else
859                         return true;
860         }
861 
862         return false;
863 }
864 
865 static void enqueue_huge_page(struct hstate *h, struct page *page)
866 {
867         int nid = page_to_nid(page);
868         list_move(&page->lru, &h->hugepage_freelists[nid]);
869         h->free_huge_pages++;
870         h->free_huge_pages_node[nid]++;
871 }
872 
873 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
874 {
875         struct page *page;
876 
877         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
878                 if (!PageHWPoison(page))
879                         break;
880         /*
881          * if 'non-isolated free hugepage' not found on the list,
882          * the allocation fails.
883          */
884         if (&h->hugepage_freelists[nid] == &page->lru)
885                 return NULL;
886         list_move(&page->lru, &h->hugepage_activelist);
887         set_page_refcounted(page);
888         h->free_huge_pages--;
889         h->free_huge_pages_node[nid]--;
890         return page;
891 }
892 
893 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
894                 nodemask_t *nmask)
895 {
896         unsigned int cpuset_mems_cookie;
897         struct zonelist *zonelist;
898         struct zone *zone;
899         struct zoneref *z;
900         int node = NUMA_NO_NODE;
901 
902         zonelist = node_zonelist(nid, gfp_mask);
903 
904 retry_cpuset:
905         cpuset_mems_cookie = read_mems_allowed_begin();
906         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
907                 struct page *page;
908 
909                 if (!cpuset_zone_allowed(zone, gfp_mask))
910                         continue;
911                 /*
912                  * no need to ask again on the same node. Pool is node rather than
913                  * zone aware
914                  */
915                 if (zone_to_nid(zone) == node)
916                         continue;
917                 node = zone_to_nid(zone);
918 
919                 page = dequeue_huge_page_node_exact(h, node);
920                 if (page)
921                         return page;
922         }
923         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
924                 goto retry_cpuset;
925 
926         return NULL;
927 }
928 
929 /* Movability of hugepages depends on migration support. */
930 static inline gfp_t htlb_alloc_mask(struct hstate *h)
931 {
932         if (hugepage_movable_supported(h))
933                 return GFP_HIGHUSER_MOVABLE;
934         else
935                 return GFP_HIGHUSER;
936 }
937 
938 static struct page *dequeue_huge_page_vma(struct hstate *h,
939                                 struct vm_area_struct *vma,
940                                 unsigned long address, int avoid_reserve,
941                                 long chg)
942 {
943         struct page *page;
944         struct mempolicy *mpol;
945         gfp_t gfp_mask;
946         nodemask_t *nodemask;
947         int nid;
948 
949         /*
950          * A child process with MAP_PRIVATE mappings created by their parent
951          * have no page reserves. This check ensures that reservations are
952          * not "stolen". The child may still get SIGKILLed
953          */
954         if (!vma_has_reserves(vma, chg) &&
955                         h->free_huge_pages - h->resv_huge_pages == 0)
956                 goto err;
957 
958         /* If reserves cannot be used, ensure enough pages are in the pool */
959         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
960                 goto err;
961 
962         gfp_mask = htlb_alloc_mask(h);
963         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
964         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
965         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
966                 SetPagePrivate(page);
967                 h->resv_huge_pages--;
968         }
969 
970         mpol_cond_put(mpol);
971         return page;
972 
973 err:
974         return NULL;
975 }
976 
977 /*
978  * common helper functions for hstate_next_node_to_{alloc|free}.
979  * We may have allocated or freed a huge page based on a different
980  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
981  * be outside of *nodes_allowed.  Ensure that we use an allowed
982  * node for alloc or free.
983  */
984 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
985 {
986         nid = next_node_in(nid, *nodes_allowed);
987         VM_BUG_ON(nid >= MAX_NUMNODES);
988 
989         return nid;
990 }
991 
992 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
993 {
994         if (!node_isset(nid, *nodes_allowed))
995                 nid = next_node_allowed(nid, nodes_allowed);
996         return nid;
997 }
998 
999 /*
1000  * returns the previously saved node ["this node"] from which to
1001  * allocate a persistent huge page for the pool and advance the
1002  * next node from which to allocate, handling wrap at end of node
1003  * mask.
1004  */
1005 static int hstate_next_node_to_alloc(struct hstate *h,
1006                                         nodemask_t *nodes_allowed)
1007 {
1008         int nid;
1009 
1010         VM_BUG_ON(!nodes_allowed);
1011 
1012         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1013         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1014 
1015         return nid;
1016 }
1017 
1018 /*
1019  * helper for free_pool_huge_page() - return the previously saved
1020  * node ["this node"] from which to free a huge page.  Advance the
1021  * next node id whether or not we find a free huge page to free so
1022  * that the next attempt to free addresses the next node.
1023  */
1024 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1025 {
1026         int nid;
1027 
1028         VM_BUG_ON(!nodes_allowed);
1029 
1030         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1031         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1032 
1033         return nid;
1034 }
1035 
1036 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1037         for (nr_nodes = nodes_weight(*mask);                            \
1038                 nr_nodes > 0 &&                                         \
1039                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1040                 nr_nodes--)
1041 
1042 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1043         for (nr_nodes = nodes_weight(*mask);                            \
1044                 nr_nodes > 0 &&                                         \
1045                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1046                 nr_nodes--)
1047 
1048 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1049 static void destroy_compound_gigantic_page(struct page *page,
1050                                         unsigned int order)
1051 {
1052         int i;
1053         int nr_pages = 1 << order;
1054         struct page *p = page + 1;
1055 
1056         atomic_set(compound_mapcount_ptr(page), 0);
1057         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1058                 clear_compound_head(p);
1059                 set_page_refcounted(p);
1060         }
1061 
1062         set_compound_order(page, 0);
1063         __ClearPageHead(page);
1064 }
1065 
1066 static void free_gigantic_page(struct page *page, unsigned int order)
1067 {
1068         free_contig_range(page_to_pfn(page), 1 << order);
1069 }
1070 
1071 #ifdef CONFIG_CONTIG_ALLOC
1072 static int __alloc_gigantic_page(unsigned long start_pfn,
1073                                 unsigned long nr_pages, gfp_t gfp_mask)
1074 {
1075         unsigned long end_pfn = start_pfn + nr_pages;
1076         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1077                                   gfp_mask);
1078 }
1079 
1080 static bool pfn_range_valid_gigantic(struct zone *z,
1081                         unsigned long start_pfn, unsigned long nr_pages)
1082 {
1083         unsigned long i, end_pfn = start_pfn + nr_pages;
1084         struct page *page;
1085 
1086         for (i = start_pfn; i < end_pfn; i++) {
1087                 if (!pfn_valid(i))
1088                         return false;
1089 
1090                 page = pfn_to_page(i);
1091 
1092                 if (page_zone(page) != z)
1093                         return false;
1094 
1095                 if (PageReserved(page))
1096                         return false;
1097 
1098                 if (page_count(page) > 0)
1099                         return false;
1100 
1101                 if (PageHuge(page))
1102                         return false;
1103         }
1104 
1105         return true;
1106 }
1107 
1108 static bool zone_spans_last_pfn(const struct zone *zone,
1109                         unsigned long start_pfn, unsigned long nr_pages)
1110 {
1111         unsigned long last_pfn = start_pfn + nr_pages - 1;
1112         return zone_spans_pfn(zone, last_pfn);
1113 }
1114 
1115 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1116                 int nid, nodemask_t *nodemask)
1117 {
1118         unsigned int order = huge_page_order(h);
1119         unsigned long nr_pages = 1 << order;
1120         unsigned long ret, pfn, flags;
1121         struct zonelist *zonelist;
1122         struct zone *zone;
1123         struct zoneref *z;
1124 
1125         zonelist = node_zonelist(nid, gfp_mask);
1126         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1127                 spin_lock_irqsave(&zone->lock, flags);
1128 
1129                 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1130                 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1131                         if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1132                                 /*
1133                                  * We release the zone lock here because
1134                                  * alloc_contig_range() will also lock the zone
1135                                  * at some point. If there's an allocation
1136                                  * spinning on this lock, it may win the race
1137                                  * and cause alloc_contig_range() to fail...
1138                                  */
1139                                 spin_unlock_irqrestore(&zone->lock, flags);
1140                                 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1141                                 if (!ret)
1142                                         return pfn_to_page(pfn);
1143                                 spin_lock_irqsave(&zone->lock, flags);
1144                         }
1145                         pfn += nr_pages;
1146                 }
1147 
1148                 spin_unlock_irqrestore(&zone->lock, flags);
1149         }
1150 
1151         return NULL;
1152 }
1153 
1154 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1155 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1156 #else /* !CONFIG_CONTIG_ALLOC */
1157 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1158                                         int nid, nodemask_t *nodemask)
1159 {
1160         return NULL;
1161 }
1162 #endif /* CONFIG_CONTIG_ALLOC */
1163 
1164 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1165 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1166                                         int nid, nodemask_t *nodemask)
1167 {
1168         return NULL;
1169 }
1170 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1171 static inline void destroy_compound_gigantic_page(struct page *page,
1172                                                 unsigned int order) { }
1173 #endif
1174 
1175 static void update_and_free_page(struct hstate *h, struct page *page)
1176 {
1177         int i;
1178 
1179         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1180                 return;
1181 
1182         h->nr_huge_pages--;
1183         h->nr_huge_pages_node[page_to_nid(page)]--;
1184         for (i = 0; i < pages_per_huge_page(h); i++) {
1185                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1186                                 1 << PG_referenced | 1 << PG_dirty |
1187                                 1 << PG_active | 1 << PG_private |
1188                                 1 << PG_writeback);
1189         }
1190         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1191         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1192         set_page_refcounted(page);
1193         if (hstate_is_gigantic(h)) {
1194                 destroy_compound_gigantic_page(page, huge_page_order(h));
1195                 free_gigantic_page(page, huge_page_order(h));
1196         } else {
1197                 __free_pages(page, huge_page_order(h));
1198         }
1199 }
1200 
1201 struct hstate *size_to_hstate(unsigned long size)
1202 {
1203         struct hstate *h;
1204 
1205         for_each_hstate(h) {
1206                 if (huge_page_size(h) == size)
1207                         return h;
1208         }
1209         return NULL;
1210 }
1211 
1212 /*
1213  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1214  * to hstate->hugepage_activelist.)
1215  *
1216  * This function can be called for tail pages, but never returns true for them.
1217  */
1218 bool page_huge_active(struct page *page)
1219 {
1220         VM_BUG_ON_PAGE(!PageHuge(page), page);
1221         return PageHead(page) && PagePrivate(&page[1]);
1222 }
1223 
1224 /* never called for tail page */
1225 static void set_page_huge_active(struct page *page)
1226 {
1227         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1228         SetPagePrivate(&page[1]);
1229 }
1230 
1231 static void clear_page_huge_active(struct page *page)
1232 {
1233         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1234         ClearPagePrivate(&page[1]);
1235 }
1236 
1237 /*
1238  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1239  * code
1240  */
1241 static inline bool PageHugeTemporary(struct page *page)
1242 {
1243         if (!PageHuge(page))
1244                 return false;
1245 
1246         return (unsigned long)page[2].mapping == -1U;
1247 }
1248 
1249 static inline void SetPageHugeTemporary(struct page *page)
1250 {
1251         page[2].mapping = (void *)-1U;
1252 }
1253 
1254 static inline void ClearPageHugeTemporary(struct page *page)
1255 {
1256         page[2].mapping = NULL;
1257 }
1258 
1259 void free_huge_page(struct page *page)
1260 {
1261         /*
1262          * Can't pass hstate in here because it is called from the
1263          * compound page destructor.
1264          */
1265         struct hstate *h = page_hstate(page);
1266         int nid = page_to_nid(page);
1267         struct hugepage_subpool *spool =
1268                 (struct hugepage_subpool *)page_private(page);
1269         bool restore_reserve;
1270 
1271         VM_BUG_ON_PAGE(page_count(page), page);
1272         VM_BUG_ON_PAGE(page_mapcount(page), page);
1273 
1274         set_page_private(page, 0);
1275         page->mapping = NULL;
1276         restore_reserve = PagePrivate(page);
1277         ClearPagePrivate(page);
1278 
1279         /*
1280          * If PagePrivate() was set on page, page allocation consumed a
1281          * reservation.  If the page was associated with a subpool, there
1282          * would have been a page reserved in the subpool before allocation
1283          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1284          * reservtion, do not call hugepage_subpool_put_pages() as this will
1285          * remove the reserved page from the subpool.
1286          */
1287         if (!restore_reserve) {
1288                 /*
1289                  * A return code of zero implies that the subpool will be
1290                  * under its minimum size if the reservation is not restored
1291                  * after page is free.  Therefore, force restore_reserve
1292                  * operation.
1293                  */
1294                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1295                         restore_reserve = true;
1296         }
1297 
1298         spin_lock(&hugetlb_lock);
1299         clear_page_huge_active(page);
1300         hugetlb_cgroup_uncharge_page(hstate_index(h),
1301                                      pages_per_huge_page(h), page);
1302         if (restore_reserve)
1303                 h->resv_huge_pages++;
1304 
1305         if (PageHugeTemporary(page)) {
1306                 list_del(&page->lru);
1307                 ClearPageHugeTemporary(page);
1308                 update_and_free_page(h, page);
1309         } else if (h->surplus_huge_pages_node[nid]) {
1310                 /* remove the page from active list */
1311                 list_del(&page->lru);
1312                 update_and_free_page(h, page);
1313                 h->surplus_huge_pages--;
1314                 h->surplus_huge_pages_node[nid]--;
1315         } else {
1316                 arch_clear_hugepage_flags(page);
1317                 enqueue_huge_page(h, page);
1318         }
1319         spin_unlock(&hugetlb_lock);
1320 }
1321 
1322 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1323 {
1324         INIT_LIST_HEAD(&page->lru);
1325         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1326         spin_lock(&hugetlb_lock);
1327         set_hugetlb_cgroup(page, NULL);
1328         h->nr_huge_pages++;
1329         h->nr_huge_pages_node[nid]++;
1330         spin_unlock(&hugetlb_lock);
1331 }
1332 
1333 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1334 {
1335         int i;
1336         int nr_pages = 1 << order;
1337         struct page *p = page + 1;
1338 
1339         /* we rely on prep_new_huge_page to set the destructor */
1340         set_compound_order(page, order);
1341         __ClearPageReserved(page);
1342         __SetPageHead(page);
1343         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1344                 /*
1345                  * For gigantic hugepages allocated through bootmem at
1346                  * boot, it's safer to be consistent with the not-gigantic
1347                  * hugepages and clear the PG_reserved bit from all tail pages
1348                  * too.  Otherwse drivers using get_user_pages() to access tail
1349                  * pages may get the reference counting wrong if they see
1350                  * PG_reserved set on a tail page (despite the head page not
1351                  * having PG_reserved set).  Enforcing this consistency between
1352                  * head and tail pages allows drivers to optimize away a check
1353                  * on the head page when they need know if put_page() is needed
1354                  * after get_user_pages().
1355                  */
1356                 __ClearPageReserved(p);
1357                 set_page_count(p, 0);
1358                 set_compound_head(p, page);
1359         }
1360         atomic_set(compound_mapcount_ptr(page), -1);
1361 }
1362 
1363 /*
1364  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1365  * transparent huge pages.  See the PageTransHuge() documentation for more
1366  * details.
1367  */
1368 int PageHuge(struct page *page)
1369 {
1370         if (!PageCompound(page))
1371                 return 0;
1372 
1373         page = compound_head(page);
1374         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1375 }
1376 EXPORT_SYMBOL_GPL(PageHuge);
1377 
1378 /*
1379  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1380  * normal or transparent huge pages.
1381  */
1382 int PageHeadHuge(struct page *page_head)
1383 {
1384         if (!PageHead(page_head))
1385                 return 0;
1386 
1387         return get_compound_page_dtor(page_head) == free_huge_page;
1388 }
1389 
1390 pgoff_t __basepage_index(struct page *page)
1391 {
1392         struct page *page_head = compound_head(page);
1393         pgoff_t index = page_index(page_head);
1394         unsigned long compound_idx;
1395 
1396         if (!PageHuge(page_head))
1397                 return page_index(page);
1398 
1399         if (compound_order(page_head) >= MAX_ORDER)
1400                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1401         else
1402                 compound_idx = page - page_head;
1403 
1404         return (index << compound_order(page_head)) + compound_idx;
1405 }
1406 
1407 static struct page *alloc_buddy_huge_page(struct hstate *h,
1408                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1409 {
1410         int order = huge_page_order(h);
1411         struct page *page;
1412 
1413         gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1414         if (nid == NUMA_NO_NODE)
1415                 nid = numa_mem_id();
1416         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1417         if (page)
1418                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1419         else
1420                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1421 
1422         return page;
1423 }
1424 
1425 /*
1426  * Common helper to allocate a fresh hugetlb page. All specific allocators
1427  * should use this function to get new hugetlb pages
1428  */
1429 static struct page *alloc_fresh_huge_page(struct hstate *h,
1430                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1431 {
1432         struct page *page;
1433 
1434         if (hstate_is_gigantic(h))
1435                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1436         else
1437                 page = alloc_buddy_huge_page(h, gfp_mask,
1438                                 nid, nmask);
1439         if (!page)
1440                 return NULL;
1441 
1442         if (hstate_is_gigantic(h))
1443                 prep_compound_gigantic_page(page, huge_page_order(h));
1444         prep_new_huge_page(h, page, page_to_nid(page));
1445 
1446         return page;
1447 }
1448 
1449 /*
1450  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1451  * manner.
1452  */
1453 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1454 {
1455         struct page *page;
1456         int nr_nodes, node;
1457         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1458 
1459         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1460                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1461                 if (page)
1462                         break;
1463         }
1464 
1465         if (!page)
1466                 return 0;
1467 
1468         put_page(page); /* free it into the hugepage allocator */
1469 
1470         return 1;
1471 }
1472 
1473 /*
1474  * Free huge page from pool from next node to free.
1475  * Attempt to keep persistent huge pages more or less
1476  * balanced over allowed nodes.
1477  * Called with hugetlb_lock locked.
1478  */
1479 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1480                                                          bool acct_surplus)
1481 {
1482         int nr_nodes, node;
1483         int ret = 0;
1484 
1485         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1486                 /*
1487                  * If we're returning unused surplus pages, only examine
1488                  * nodes with surplus pages.
1489                  */
1490                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1491                     !list_empty(&h->hugepage_freelists[node])) {
1492                         struct page *page =
1493                                 list_entry(h->hugepage_freelists[node].next,
1494                                           struct page, lru);
1495                         list_del(&page->lru);
1496                         h->free_huge_pages--;
1497                         h->free_huge_pages_node[node]--;
1498                         if (acct_surplus) {
1499                                 h->surplus_huge_pages--;
1500                                 h->surplus_huge_pages_node[node]--;
1501                         }
1502                         update_and_free_page(h, page);
1503                         ret = 1;
1504                         break;
1505                 }
1506         }
1507 
1508         return ret;
1509 }
1510 
1511 /*
1512  * Dissolve a given free hugepage into free buddy pages. This function does
1513  * nothing for in-use hugepages and non-hugepages.
1514  * This function returns values like below:
1515  *
1516  *  -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1517  *          (allocated or reserved.)
1518  *       0: successfully dissolved free hugepages or the page is not a
1519  *          hugepage (considered as already dissolved)
1520  */
1521 int dissolve_free_huge_page(struct page *page)
1522 {
1523         int rc = -EBUSY;
1524 
1525         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1526         if (!PageHuge(page))
1527                 return 0;
1528 
1529         spin_lock(&hugetlb_lock);
1530         if (!PageHuge(page)) {
1531                 rc = 0;
1532                 goto out;
1533         }
1534 
1535         if (!page_count(page)) {
1536                 struct page *head = compound_head(page);
1537                 struct hstate *h = page_hstate(head);
1538                 int nid = page_to_nid(head);
1539                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1540                         goto out;
1541                 /*
1542                  * Move PageHWPoison flag from head page to the raw error page,
1543                  * which makes any subpages rather than the error page reusable.
1544                  */
1545                 if (PageHWPoison(head) && page != head) {
1546                         SetPageHWPoison(page);
1547                         ClearPageHWPoison(head);
1548                 }
1549                 list_del(&head->lru);
1550                 h->free_huge_pages--;
1551                 h->free_huge_pages_node[nid]--;
1552                 h->max_huge_pages--;
1553                 update_and_free_page(h, head);
1554                 rc = 0;
1555         }
1556 out:
1557         spin_unlock(&hugetlb_lock);
1558         return rc;
1559 }
1560 
1561 /*
1562  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1563  * make specified memory blocks removable from the system.
1564  * Note that this will dissolve a free gigantic hugepage completely, if any
1565  * part of it lies within the given range.
1566  * Also note that if dissolve_free_huge_page() returns with an error, all
1567  * free hugepages that were dissolved before that error are lost.
1568  */
1569 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1570 {
1571         unsigned long pfn;
1572         struct page *page;
1573         int rc = 0;
1574 
1575         if (!hugepages_supported())
1576                 return rc;
1577 
1578         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1579                 page = pfn_to_page(pfn);
1580                 rc = dissolve_free_huge_page(page);
1581                 if (rc)
1582                         break;
1583         }
1584 
1585         return rc;
1586 }
1587 
1588 /*
1589  * Allocates a fresh surplus page from the page allocator.
1590  */
1591 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1592                 int nid, nodemask_t *nmask)
1593 {
1594         struct page *page = NULL;
1595 
1596         if (hstate_is_gigantic(h))
1597                 return NULL;
1598 
1599         spin_lock(&hugetlb_lock);
1600         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1601                 goto out_unlock;
1602         spin_unlock(&hugetlb_lock);
1603 
1604         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1605         if (!page)
1606                 return NULL;
1607 
1608         spin_lock(&hugetlb_lock);
1609         /*
1610          * We could have raced with the pool size change.
1611          * Double check that and simply deallocate the new page
1612          * if we would end up overcommiting the surpluses. Abuse
1613          * temporary page to workaround the nasty free_huge_page
1614          * codeflow
1615          */
1616         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1617                 SetPageHugeTemporary(page);
1618                 spin_unlock(&hugetlb_lock);
1619                 put_page(page);
1620                 return NULL;
1621         } else {
1622                 h->surplus_huge_pages++;
1623                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1624         }
1625 
1626 out_unlock:
1627         spin_unlock(&hugetlb_lock);
1628 
1629         return page;
1630 }
1631 
1632 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1633                                      int nid, nodemask_t *nmask)
1634 {
1635         struct page *page;
1636 
1637         if (hstate_is_gigantic(h))
1638                 return NULL;
1639 
1640         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1641         if (!page)
1642                 return NULL;
1643 
1644         /*
1645          * We do not account these pages as surplus because they are only
1646          * temporary and will be released properly on the last reference
1647          */
1648         SetPageHugeTemporary(page);
1649 
1650         return page;
1651 }
1652 
1653 /*
1654  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1655  */
1656 static
1657 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1658                 struct vm_area_struct *vma, unsigned long addr)
1659 {
1660         struct page *page;
1661         struct mempolicy *mpol;
1662         gfp_t gfp_mask = htlb_alloc_mask(h);
1663         int nid;
1664         nodemask_t *nodemask;
1665 
1666         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1667         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1668         mpol_cond_put(mpol);
1669 
1670         return page;
1671 }
1672 
1673 /* page migration callback function */
1674 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1675 {
1676         gfp_t gfp_mask = htlb_alloc_mask(h);
1677         struct page *page = NULL;
1678 
1679         if (nid != NUMA_NO_NODE)
1680                 gfp_mask |= __GFP_THISNODE;
1681 
1682         spin_lock(&hugetlb_lock);
1683         if (h->free_huge_pages - h->resv_huge_pages > 0)
1684                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1685         spin_unlock(&hugetlb_lock);
1686 
1687         if (!page)
1688                 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1689 
1690         return page;
1691 }
1692 
1693 /* page migration callback function */
1694 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1695                 nodemask_t *nmask)
1696 {
1697         gfp_t gfp_mask = htlb_alloc_mask(h);
1698 
1699         spin_lock(&hugetlb_lock);
1700         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1701                 struct page *page;
1702 
1703                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1704                 if (page) {
1705                         spin_unlock(&hugetlb_lock);
1706                         return page;
1707                 }
1708         }
1709         spin_unlock(&hugetlb_lock);
1710 
1711         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1712 }
1713 
1714 /* mempolicy aware migration callback */
1715 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1716                 unsigned long address)
1717 {
1718         struct mempolicy *mpol;
1719         nodemask_t *nodemask;
1720         struct page *page;
1721         gfp_t gfp_mask;
1722         int node;
1723 
1724         gfp_mask = htlb_alloc_mask(h);
1725         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1726         page = alloc_huge_page_nodemask(h, node, nodemask);
1727         mpol_cond_put(mpol);
1728 
1729         return page;
1730 }
1731 
1732 /*
1733  * Increase the hugetlb pool such that it can accommodate a reservation
1734  * of size 'delta'.
1735  */
1736 static int gather_surplus_pages(struct hstate *h, int delta)
1737 {
1738         struct list_head surplus_list;
1739         struct page *page, *tmp;
1740         int ret, i;
1741         int needed, allocated;
1742         bool alloc_ok = true;
1743 
1744         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1745         if (needed <= 0) {
1746                 h->resv_huge_pages += delta;
1747                 return 0;
1748         }
1749 
1750         allocated = 0;
1751         INIT_LIST_HEAD(&surplus_list);
1752 
1753         ret = -ENOMEM;
1754 retry:
1755         spin_unlock(&hugetlb_lock);
1756         for (i = 0; i < needed; i++) {
1757                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1758                                 NUMA_NO_NODE, NULL);
1759                 if (!page) {
1760                         alloc_ok = false;
1761                         break;
1762                 }
1763                 list_add(&page->lru, &surplus_list);
1764                 cond_resched();
1765         }
1766         allocated += i;
1767 
1768         /*
1769          * After retaking hugetlb_lock, we need to recalculate 'needed'
1770          * because either resv_huge_pages or free_huge_pages may have changed.
1771          */
1772         spin_lock(&hugetlb_lock);
1773         needed = (h->resv_huge_pages + delta) -
1774                         (h->free_huge_pages + allocated);
1775         if (needed > 0) {
1776                 if (alloc_ok)
1777                         goto retry;
1778                 /*
1779                  * We were not able to allocate enough pages to
1780                  * satisfy the entire reservation so we free what
1781                  * we've allocated so far.
1782                  */
1783                 goto free;
1784         }
1785         /*
1786          * The surplus_list now contains _at_least_ the number of extra pages
1787          * needed to accommodate the reservation.  Add the appropriate number
1788          * of pages to the hugetlb pool and free the extras back to the buddy
1789          * allocator.  Commit the entire reservation here to prevent another
1790          * process from stealing the pages as they are added to the pool but
1791          * before they are reserved.
1792          */
1793         needed += allocated;
1794         h->resv_huge_pages += delta;
1795         ret = 0;
1796 
1797         /* Free the needed pages to the hugetlb pool */
1798         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1799                 if ((--needed) < 0)
1800                         break;
1801                 /*
1802                  * This page is now managed by the hugetlb allocator and has
1803                  * no users -- drop the buddy allocator's reference.
1804                  */
1805                 put_page_testzero(page);
1806                 VM_BUG_ON_PAGE(page_count(page), page);
1807                 enqueue_huge_page(h, page);
1808         }
1809 free:
1810         spin_unlock(&hugetlb_lock);
1811 
1812         /* Free unnecessary surplus pages to the buddy allocator */
1813         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1814                 put_page(page);
1815         spin_lock(&hugetlb_lock);
1816 
1817         return ret;
1818 }
1819 
1820 /*
1821  * This routine has two main purposes:
1822  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1823  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1824  *    to the associated reservation map.
1825  * 2) Free any unused surplus pages that may have been allocated to satisfy
1826  *    the reservation.  As many as unused_resv_pages may be freed.
1827  *
1828  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1829  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1830  * we must make sure nobody else can claim pages we are in the process of
1831  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1832  * number of huge pages we plan to free when dropping the lock.
1833  */
1834 static void return_unused_surplus_pages(struct hstate *h,
1835                                         unsigned long unused_resv_pages)
1836 {
1837         unsigned long nr_pages;
1838 
1839         /* Cannot return gigantic pages currently */
1840         if (hstate_is_gigantic(h))
1841                 goto out;
1842 
1843         /*
1844          * Part (or even all) of the reservation could have been backed
1845          * by pre-allocated pages. Only free surplus pages.
1846          */
1847         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1848 
1849         /*
1850          * We want to release as many surplus pages as possible, spread
1851          * evenly across all nodes with memory. Iterate across these nodes
1852          * until we can no longer free unreserved surplus pages. This occurs
1853          * when the nodes with surplus pages have no free pages.
1854          * free_pool_huge_page() will balance the the freed pages across the
1855          * on-line nodes with memory and will handle the hstate accounting.
1856          *
1857          * Note that we decrement resv_huge_pages as we free the pages.  If
1858          * we drop the lock, resv_huge_pages will still be sufficiently large
1859          * to cover subsequent pages we may free.
1860          */
1861         while (nr_pages--) {
1862                 h->resv_huge_pages--;
1863                 unused_resv_pages--;
1864                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1865                         goto out;
1866                 cond_resched_lock(&hugetlb_lock);
1867         }
1868 
1869 out:
1870         /* Fully uncommit the reservation */
1871         h->resv_huge_pages -= unused_resv_pages;
1872 }
1873 
1874 
1875 /*
1876  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1877  * are used by the huge page allocation routines to manage reservations.
1878  *
1879  * vma_needs_reservation is called to determine if the huge page at addr
1880  * within the vma has an associated reservation.  If a reservation is
1881  * needed, the value 1 is returned.  The caller is then responsible for
1882  * managing the global reservation and subpool usage counts.  After
1883  * the huge page has been allocated, vma_commit_reservation is called
1884  * to add the page to the reservation map.  If the page allocation fails,
1885  * the reservation must be ended instead of committed.  vma_end_reservation
1886  * is called in such cases.
1887  *
1888  * In the normal case, vma_commit_reservation returns the same value
1889  * as the preceding vma_needs_reservation call.  The only time this
1890  * is not the case is if a reserve map was changed between calls.  It
1891  * is the responsibility of the caller to notice the difference and
1892  * take appropriate action.
1893  *
1894  * vma_add_reservation is used in error paths where a reservation must
1895  * be restored when a newly allocated huge page must be freed.  It is
1896  * to be called after calling vma_needs_reservation to determine if a
1897  * reservation exists.
1898  */
1899 enum vma_resv_mode {
1900         VMA_NEEDS_RESV,
1901         VMA_COMMIT_RESV,
1902         VMA_END_RESV,
1903         VMA_ADD_RESV,
1904 };
1905 static long __vma_reservation_common(struct hstate *h,
1906                                 struct vm_area_struct *vma, unsigned long addr,
1907                                 enum vma_resv_mode mode)
1908 {
1909         struct resv_map *resv;
1910         pgoff_t idx;
1911         long ret;
1912 
1913         resv = vma_resv_map(vma);
1914         if (!resv)
1915                 return 1;
1916 
1917         idx = vma_hugecache_offset(h, vma, addr);
1918         switch (mode) {
1919         case VMA_NEEDS_RESV:
1920                 ret = region_chg(resv, idx, idx + 1);
1921                 break;
1922         case VMA_COMMIT_RESV:
1923                 ret = region_add(resv, idx, idx + 1);
1924                 break;
1925         case VMA_END_RESV:
1926                 region_abort(resv, idx, idx + 1);
1927                 ret = 0;
1928                 break;
1929         case VMA_ADD_RESV:
1930                 if (vma->vm_flags & VM_MAYSHARE)
1931                         ret = region_add(resv, idx, idx + 1);
1932                 else {
1933                         region_abort(resv, idx, idx + 1);
1934                         ret = region_del(resv, idx, idx + 1);
1935                 }
1936                 break;
1937         default:
1938                 BUG();
1939         }
1940 
1941         if (vma->vm_flags & VM_MAYSHARE)
1942                 return ret;
1943         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1944                 /*
1945                  * In most cases, reserves always exist for private mappings.
1946                  * However, a file associated with mapping could have been
1947                  * hole punched or truncated after reserves were consumed.
1948                  * As subsequent fault on such a range will not use reserves.
1949                  * Subtle - The reserve map for private mappings has the
1950                  * opposite meaning than that of shared mappings.  If NO
1951                  * entry is in the reserve map, it means a reservation exists.
1952                  * If an entry exists in the reserve map, it means the
1953                  * reservation has already been consumed.  As a result, the
1954                  * return value of this routine is the opposite of the
1955                  * value returned from reserve map manipulation routines above.
1956                  */
1957                 if (ret)
1958                         return 0;
1959                 else
1960                         return 1;
1961         }
1962         else
1963                 return ret < 0 ? ret : 0;
1964 }
1965 
1966 static long vma_needs_reservation(struct hstate *h,
1967                         struct vm_area_struct *vma, unsigned long addr)
1968 {
1969         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1970 }
1971 
1972 static long vma_commit_reservation(struct hstate *h,
1973                         struct vm_area_struct *vma, unsigned long addr)
1974 {
1975         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1976 }
1977 
1978 static void vma_end_reservation(struct hstate *h,
1979                         struct vm_area_struct *vma, unsigned long addr)
1980 {
1981         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1982 }
1983 
1984 static long vma_add_reservation(struct hstate *h,
1985                         struct vm_area_struct *vma, unsigned long addr)
1986 {
1987         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1988 }
1989 
1990 /*
1991  * This routine is called to restore a reservation on error paths.  In the
1992  * specific error paths, a huge page was allocated (via alloc_huge_page)
1993  * and is about to be freed.  If a reservation for the page existed,
1994  * alloc_huge_page would have consumed the reservation and set PagePrivate
1995  * in the newly allocated page.  When the page is freed via free_huge_page,
1996  * the global reservation count will be incremented if PagePrivate is set.
1997  * However, free_huge_page can not adjust the reserve map.  Adjust the
1998  * reserve map here to be consistent with global reserve count adjustments
1999  * to be made by free_huge_page.
2000  */
2001 static void restore_reserve_on_error(struct hstate *h,
2002                         struct vm_area_struct *vma, unsigned long address,
2003                         struct page *page)
2004 {
2005         if (unlikely(PagePrivate(page))) {
2006                 long rc = vma_needs_reservation(h, vma, address);
2007 
2008                 if (unlikely(rc < 0)) {
2009                         /*
2010                          * Rare out of memory condition in reserve map
2011                          * manipulation.  Clear PagePrivate so that
2012                          * global reserve count will not be incremented
2013                          * by free_huge_page.  This will make it appear
2014                          * as though the reservation for this page was
2015                          * consumed.  This may prevent the task from
2016                          * faulting in the page at a later time.  This
2017                          * is better than inconsistent global huge page
2018                          * accounting of reserve counts.
2019                          */
2020                         ClearPagePrivate(page);
2021                 } else if (rc) {
2022                         rc = vma_add_reservation(h, vma, address);
2023                         if (unlikely(rc < 0))
2024                                 /*
2025                                  * See above comment about rare out of
2026                                  * memory condition.
2027                                  */
2028                                 ClearPagePrivate(page);
2029                 } else
2030                         vma_end_reservation(h, vma, address);
2031         }
2032 }
2033 
2034 struct page *alloc_huge_page(struct vm_area_struct *vma,
2035                                     unsigned long addr, int avoid_reserve)
2036 {
2037         struct hugepage_subpool *spool = subpool_vma(vma);
2038         struct hstate *h = hstate_vma(vma);
2039         struct page *page;
2040         long map_chg, map_commit;
2041         long gbl_chg;
2042         int ret, idx;
2043         struct hugetlb_cgroup *h_cg;
2044 
2045         idx = hstate_index(h);
2046         /*
2047          * Examine the region/reserve map to determine if the process
2048          * has a reservation for the page to be allocated.  A return
2049          * code of zero indicates a reservation exists (no change).
2050          */
2051         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2052         if (map_chg < 0)
2053                 return ERR_PTR(-ENOMEM);
2054 
2055         /*
2056          * Processes that did not create the mapping will have no
2057          * reserves as indicated by the region/reserve map. Check
2058          * that the allocation will not exceed the subpool limit.
2059          * Allocations for MAP_NORESERVE mappings also need to be
2060          * checked against any subpool limit.
2061          */
2062         if (map_chg || avoid_reserve) {
2063                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2064                 if (gbl_chg < 0) {
2065                         vma_end_reservation(h, vma, addr);
2066                         return ERR_PTR(-ENOSPC);
2067                 }
2068 
2069                 /*
2070                  * Even though there was no reservation in the region/reserve
2071                  * map, there could be reservations associated with the
2072                  * subpool that can be used.  This would be indicated if the
2073                  * return value of hugepage_subpool_get_pages() is zero.
2074                  * However, if avoid_reserve is specified we still avoid even
2075                  * the subpool reservations.
2076                  */
2077                 if (avoid_reserve)
2078                         gbl_chg = 1;
2079         }
2080 
2081         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2082         if (ret)
2083                 goto out_subpool_put;
2084 
2085         spin_lock(&hugetlb_lock);
2086         /*
2087          * glb_chg is passed to indicate whether or not a page must be taken
2088          * from the global free pool (global change).  gbl_chg == 0 indicates
2089          * a reservation exists for the allocation.
2090          */
2091         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2092         if (!page) {
2093                 spin_unlock(&hugetlb_lock);
2094                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2095                 if (!page)
2096                         goto out_uncharge_cgroup;
2097                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2098                         SetPagePrivate(page);
2099                         h->resv_huge_pages--;
2100                 }
2101                 spin_lock(&hugetlb_lock);
2102                 list_move(&page->lru, &h->hugepage_activelist);
2103                 /* Fall through */
2104         }
2105         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2106         spin_unlock(&hugetlb_lock);
2107 
2108         set_page_private(page, (unsigned long)spool);
2109 
2110         map_commit = vma_commit_reservation(h, vma, addr);
2111         if (unlikely(map_chg > map_commit)) {
2112                 /*
2113                  * The page was added to the reservation map between
2114                  * vma_needs_reservation and vma_commit_reservation.
2115                  * This indicates a race with hugetlb_reserve_pages.
2116                  * Adjust for the subpool count incremented above AND
2117                  * in hugetlb_reserve_pages for the same page.  Also,
2118                  * the reservation count added in hugetlb_reserve_pages
2119                  * no longer applies.
2120                  */
2121                 long rsv_adjust;
2122 
2123                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2124                 hugetlb_acct_memory(h, -rsv_adjust);
2125         }
2126         return page;
2127 
2128 out_uncharge_cgroup:
2129         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2130 out_subpool_put:
2131         if (map_chg || avoid_reserve)
2132                 hugepage_subpool_put_pages(spool, 1);
2133         vma_end_reservation(h, vma, addr);
2134         return ERR_PTR(-ENOSPC);
2135 }
2136 
2137 int alloc_bootmem_huge_page(struct hstate *h)
2138         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2139 int __alloc_bootmem_huge_page(struct hstate *h)
2140 {
2141         struct huge_bootmem_page *m;
2142         int nr_nodes, node;
2143 
2144         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2145                 void *addr;
2146 
2147                 addr = memblock_alloc_try_nid_raw(
2148                                 huge_page_size(h), huge_page_size(h),
2149                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2150                 if (addr) {
2151                         /*
2152                          * Use the beginning of the huge page to store the
2153                          * huge_bootmem_page struct (until gather_bootmem
2154                          * puts them into the mem_map).
2155                          */
2156                         m = addr;
2157                         goto found;
2158                 }
2159         }
2160         return 0;
2161 
2162 found:
2163         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2164         /* Put them into a private list first because mem_map is not up yet */
2165         INIT_LIST_HEAD(&m->list);
2166         list_add(&m->list, &huge_boot_pages);
2167         m->hstate = h;
2168         return 1;
2169 }
2170 
2171 static void __init prep_compound_huge_page(struct page *page,
2172                 unsigned int order)
2173 {
2174         if (unlikely(order > (MAX_ORDER - 1)))
2175                 prep_compound_gigantic_page(page, order);
2176         else
2177                 prep_compound_page(page, order);
2178 }
2179 
2180 /* Put bootmem huge pages into the standard lists after mem_map is up */
2181 static void __init gather_bootmem_prealloc(void)
2182 {
2183         struct huge_bootmem_page *m;
2184 
2185         list_for_each_entry(m, &huge_boot_pages, list) {
2186                 struct page *page = virt_to_page(m);
2187                 struct hstate *h = m->hstate;
2188 
2189                 WARN_ON(page_count(page) != 1);
2190                 prep_compound_huge_page(page, h->order);
2191                 WARN_ON(PageReserved(page));
2192                 prep_new_huge_page(h, page, page_to_nid(page));
2193                 put_page(page); /* free it into the hugepage allocator */
2194 
2195                 /*
2196                  * If we had gigantic hugepages allocated at boot time, we need
2197                  * to restore the 'stolen' pages to totalram_pages in order to
2198                  * fix confusing memory reports from free(1) and another
2199                  * side-effects, like CommitLimit going negative.
2200                  */
2201                 if (hstate_is_gigantic(h))
2202                         adjust_managed_page_count(page, 1 << h->order);
2203                 cond_resched();
2204         }
2205 }
2206 
2207 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2208 {
2209         unsigned long i;
2210 
2211         for (i = 0; i < h->max_huge_pages; ++i) {
2212                 if (hstate_is_gigantic(h)) {
2213                         if (!alloc_bootmem_huge_page(h))
2214                                 break;
2215                 } else if (!alloc_pool_huge_page(h,
2216                                          &node_states[N_MEMORY]))
2217                         break;
2218                 cond_resched();
2219         }
2220         if (i < h->max_huge_pages) {
2221                 char buf[32];
2222 
2223                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2224                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2225                         h->max_huge_pages, buf, i);
2226                 h->max_huge_pages = i;
2227         }
2228 }
2229 
2230 static void __init hugetlb_init_hstates(void)
2231 {
2232         struct hstate *h;
2233 
2234         for_each_hstate(h) {
2235                 if (minimum_order > huge_page_order(h))
2236                         minimum_order = huge_page_order(h);
2237 
2238                 /* oversize hugepages were init'ed in early boot */
2239                 if (!hstate_is_gigantic(h))
2240                         hugetlb_hstate_alloc_pages(h);
2241         }
2242         VM_BUG_ON(minimum_order == UINT_MAX);
2243 }
2244 
2245 static void __init report_hugepages(void)
2246 {
2247         struct hstate *h;
2248 
2249         for_each_hstate(h) {
2250                 char buf[32];
2251 
2252                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2253                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2254                         buf, h->free_huge_pages);
2255         }
2256 }
2257 
2258 #ifdef CONFIG_HIGHMEM
2259 static void try_to_free_low(struct hstate *h, unsigned long count,
2260                                                 nodemask_t *nodes_allowed)
2261 {
2262         int i;
2263 
2264         if (hstate_is_gigantic(h))
2265                 return;
2266 
2267         for_each_node_mask(i, *nodes_allowed) {
2268                 struct page *page, *next;
2269                 struct list_head *freel = &h->hugepage_freelists[i];
2270                 list_for_each_entry_safe(page, next, freel, lru) {
2271                         if (count >= h->nr_huge_pages)
2272                                 return;
2273                         if (PageHighMem(page))
2274                                 continue;
2275                         list_del(&page->lru);
2276                         update_and_free_page(h, page);
2277                         h->free_huge_pages--;
2278                         h->free_huge_pages_node[page_to_nid(page)]--;
2279                 }
2280         }
2281 }
2282 #else
2283 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2284                                                 nodemask_t *nodes_allowed)
2285 {
2286 }
2287 #endif
2288 
2289 /*
2290  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2291  * balanced by operating on them in a round-robin fashion.
2292  * Returns 1 if an adjustment was made.
2293  */
2294 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2295                                 int delta)
2296 {
2297         int nr_nodes, node;
2298 
2299         VM_BUG_ON(delta != -1 && delta != 1);
2300 
2301         if (delta < 0) {
2302                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2303                         if (h->surplus_huge_pages_node[node])
2304                                 goto found;
2305                 }
2306         } else {
2307                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2308                         if (h->surplus_huge_pages_node[node] <
2309                                         h->nr_huge_pages_node[node])
2310                                 goto found;
2311                 }
2312         }
2313         return 0;
2314 
2315 found:
2316         h->surplus_huge_pages += delta;
2317         h->surplus_huge_pages_node[node] += delta;
2318         return 1;
2319 }
2320 
2321 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2322 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2323                               nodemask_t *nodes_allowed)
2324 {
2325         unsigned long min_count, ret;
2326 
2327         spin_lock(&hugetlb_lock);
2328 
2329         /*
2330          * Check for a node specific request.
2331          * Changing node specific huge page count may require a corresponding
2332          * change to the global count.  In any case, the passed node mask
2333          * (nodes_allowed) will restrict alloc/free to the specified node.
2334          */
2335         if (nid != NUMA_NO_NODE) {
2336                 unsigned long old_count = count;
2337 
2338                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2339                 /*
2340                  * User may have specified a large count value which caused the
2341                  * above calculation to overflow.  In this case, they wanted
2342                  * to allocate as many huge pages as possible.  Set count to
2343                  * largest possible value to align with their intention.
2344                  */
2345                 if (count < old_count)
2346                         count = ULONG_MAX;
2347         }
2348 
2349         /*
2350          * Gigantic pages runtime allocation depend on the capability for large
2351          * page range allocation.
2352          * If the system does not provide this feature, return an error when
2353          * the user tries to allocate gigantic pages but let the user free the
2354          * boottime allocated gigantic pages.
2355          */
2356         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2357                 if (count > persistent_huge_pages(h)) {
2358                         spin_unlock(&hugetlb_lock);
2359                         return -EINVAL;
2360                 }
2361                 /* Fall through to decrease pool */
2362         }
2363 
2364         /*
2365          * Increase the pool size
2366          * First take pages out of surplus state.  Then make up the
2367          * remaining difference by allocating fresh huge pages.
2368          *
2369          * We might race with alloc_surplus_huge_page() here and be unable
2370          * to convert a surplus huge page to a normal huge page. That is
2371          * not critical, though, it just means the overall size of the
2372          * pool might be one hugepage larger than it needs to be, but
2373          * within all the constraints specified by the sysctls.
2374          */
2375         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2376                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2377                         break;
2378         }
2379 
2380         while (count > persistent_huge_pages(h)) {
2381                 /*
2382                  * If this allocation races such that we no longer need the
2383                  * page, free_huge_page will handle it by freeing the page
2384                  * and reducing the surplus.
2385                  */
2386                 spin_unlock(&hugetlb_lock);
2387 
2388                 /* yield cpu to avoid soft lockup */
2389                 cond_resched();
2390 
2391                 ret = alloc_pool_huge_page(h, nodes_allowed);
2392                 spin_lock(&hugetlb_lock);
2393                 if (!ret)
2394                         goto out;
2395 
2396                 /* Bail for signals. Probably ctrl-c from user */
2397                 if (signal_pending(current))
2398                         goto out;
2399         }
2400 
2401         /*
2402          * Decrease the pool size
2403          * First return free pages to the buddy allocator (being careful
2404          * to keep enough around to satisfy reservations).  Then place
2405          * pages into surplus state as needed so the pool will shrink
2406          * to the desired size as pages become free.
2407          *
2408          * By placing pages into the surplus state independent of the
2409          * overcommit value, we are allowing the surplus pool size to
2410          * exceed overcommit. There are few sane options here. Since
2411          * alloc_surplus_huge_page() is checking the global counter,
2412          * though, we'll note that we're not allowed to exceed surplus
2413          * and won't grow the pool anywhere else. Not until one of the
2414          * sysctls are changed, or the surplus pages go out of use.
2415          */
2416         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2417         min_count = max(count, min_count);
2418         try_to_free_low(h, min_count, nodes_allowed);
2419         while (min_count < persistent_huge_pages(h)) {
2420                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2421                         break;
2422                 cond_resched_lock(&hugetlb_lock);
2423         }
2424         while (count < persistent_huge_pages(h)) {
2425                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2426                         break;
2427         }
2428 out:
2429         h->max_huge_pages = persistent_huge_pages(h);
2430         spin_unlock(&hugetlb_lock);
2431 
2432         return 0;
2433 }
2434 
2435 #define HSTATE_ATTR_RO(_name) \
2436         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2437 
2438 #define HSTATE_ATTR(_name) \
2439         static struct kobj_attribute _name##_attr = \
2440                 __ATTR(_name, 0644, _name##_show, _name##_store)
2441 
2442 static struct kobject *hugepages_kobj;
2443 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2444 
2445 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2446 
2447 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2448 {
2449         int i;
2450 
2451         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2452                 if (hstate_kobjs[i] == kobj) {
2453                         if (nidp)
2454                                 *nidp = NUMA_NO_NODE;
2455                         return &hstates[i];
2456                 }
2457 
2458         return kobj_to_node_hstate(kobj, nidp);
2459 }
2460 
2461 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2462                                         struct kobj_attribute *attr, char *buf)
2463 {
2464         struct hstate *h;
2465         unsigned long nr_huge_pages;
2466         int nid;
2467 
2468         h = kobj_to_hstate(kobj, &nid);
2469         if (nid == NUMA_NO_NODE)
2470                 nr_huge_pages = h->nr_huge_pages;
2471         else
2472                 nr_huge_pages = h->nr_huge_pages_node[nid];
2473 
2474         return sprintf(buf, "%lu\n", nr_huge_pages);
2475 }
2476 
2477 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2478                                            struct hstate *h, int nid,
2479                                            unsigned long count, size_t len)
2480 {
2481         int err;
2482         nodemask_t nodes_allowed, *n_mask;
2483 
2484         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2485                 return -EINVAL;
2486 
2487         if (nid == NUMA_NO_NODE) {
2488                 /*
2489                  * global hstate attribute
2490                  */
2491                 if (!(obey_mempolicy &&
2492                                 init_nodemask_of_mempolicy(&nodes_allowed)))
2493                         n_mask = &node_states[N_MEMORY];
2494                 else
2495                         n_mask = &nodes_allowed;
2496         } else {
2497                 /*
2498                  * Node specific request.  count adjustment happens in
2499                  * set_max_huge_pages() after acquiring hugetlb_lock.
2500                  */
2501                 init_nodemask_of_node(&nodes_allowed, nid);
2502                 n_mask = &nodes_allowed;
2503         }
2504 
2505         err = set_max_huge_pages(h, count, nid, n_mask);
2506 
2507         return err ? err : len;
2508 }
2509 
2510 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2511                                          struct kobject *kobj, const char *buf,
2512                                          size_t len)
2513 {
2514         struct hstate *h;
2515         unsigned long count;
2516         int nid;
2517         int err;
2518 
2519         err = kstrtoul(buf, 10, &count);
2520         if (err)
2521                 return err;
2522 
2523         h = kobj_to_hstate(kobj, &nid);
2524         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2525 }
2526 
2527 static ssize_t nr_hugepages_show(struct kobject *kobj,
2528                                        struct kobj_attribute *attr, char *buf)
2529 {
2530         return nr_hugepages_show_common(kobj, attr, buf);
2531 }
2532 
2533 static ssize_t nr_hugepages_store(struct kobject *kobj,
2534                struct kobj_attribute *attr, const char *buf, size_t len)
2535 {
2536         return nr_hugepages_store_common(false, kobj, buf, len);
2537 }
2538 HSTATE_ATTR(nr_hugepages);
2539 
2540 #ifdef CONFIG_NUMA
2541 
2542 /*
2543  * hstate attribute for optionally mempolicy-based constraint on persistent
2544  * huge page alloc/free.
2545  */
2546 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2547                                        struct kobj_attribute *attr, char *buf)
2548 {
2549         return nr_hugepages_show_common(kobj, attr, buf);
2550 }
2551 
2552 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2553                struct kobj_attribute *attr, const char *buf, size_t len)
2554 {
2555         return nr_hugepages_store_common(true, kobj, buf, len);
2556 }
2557 HSTATE_ATTR(nr_hugepages_mempolicy);
2558 #endif
2559 
2560 
2561 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2562                                         struct kobj_attribute *attr, char *buf)
2563 {
2564         struct hstate *h = kobj_to_hstate(kobj, NULL);
2565         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2566 }
2567 
2568 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2569                 struct kobj_attribute *attr, const char *buf, size_t count)
2570 {
2571         int err;
2572         unsigned long input;
2573         struct hstate *h = kobj_to_hstate(kobj, NULL);
2574 
2575         if (hstate_is_gigantic(h))
2576                 return -EINVAL;
2577 
2578         err = kstrtoul(buf, 10, &input);
2579         if (err)
2580                 return err;
2581 
2582         spin_lock(&hugetlb_lock);
2583         h->nr_overcommit_huge_pages = input;
2584         spin_unlock(&hugetlb_lock);
2585 
2586         return count;
2587 }
2588 HSTATE_ATTR(nr_overcommit_hugepages);
2589 
2590 static ssize_t free_hugepages_show(struct kobject *kobj,
2591                                         struct kobj_attribute *attr, char *buf)
2592 {
2593         struct hstate *h;
2594         unsigned long free_huge_pages;
2595         int nid;
2596 
2597         h = kobj_to_hstate(kobj, &nid);
2598         if (nid == NUMA_NO_NODE)
2599                 free_huge_pages = h->free_huge_pages;
2600         else
2601                 free_huge_pages = h->free_huge_pages_node[nid];
2602 
2603         return sprintf(buf, "%lu\n", free_huge_pages);
2604 }
2605 HSTATE_ATTR_RO(free_hugepages);
2606 
2607 static ssize_t resv_hugepages_show(struct kobject *kobj,
2608                                         struct kobj_attribute *attr, char *buf)
2609 {
2610         struct hstate *h = kobj_to_hstate(kobj, NULL);
2611         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2612 }
2613 HSTATE_ATTR_RO(resv_hugepages);
2614 
2615 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2616                                         struct kobj_attribute *attr, char *buf)
2617 {
2618         struct hstate *h;
2619         unsigned long surplus_huge_pages;
2620         int nid;
2621 
2622         h = kobj_to_hstate(kobj, &nid);
2623         if (nid == NUMA_NO_NODE)
2624                 surplus_huge_pages = h->surplus_huge_pages;
2625         else
2626                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2627 
2628         return sprintf(buf, "%lu\n", surplus_huge_pages);
2629 }
2630 HSTATE_ATTR_RO(surplus_hugepages);
2631 
2632 static struct attribute *hstate_attrs[] = {
2633         &nr_hugepages_attr.attr,
2634         &nr_overcommit_hugepages_attr.attr,
2635         &free_hugepages_attr.attr,
2636         &resv_hugepages_attr.attr,
2637         &surplus_hugepages_attr.attr,
2638 #ifdef CONFIG_NUMA
2639         &nr_hugepages_mempolicy_attr.attr,
2640 #endif
2641         NULL,
2642 };
2643 
2644 static const struct attribute_group hstate_attr_group = {
2645         .attrs = hstate_attrs,
2646 };
2647 
2648 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2649                                     struct kobject **hstate_kobjs,
2650                                     const struct attribute_group *hstate_attr_group)
2651 {
2652         int retval;
2653         int hi = hstate_index(h);
2654 
2655         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2656         if (!hstate_kobjs[hi])
2657                 return -ENOMEM;
2658 
2659         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2660         if (retval)
2661                 kobject_put(hstate_kobjs[hi]);
2662 
2663         return retval;
2664 }
2665 
2666 static void __init hugetlb_sysfs_init(void)
2667 {
2668         struct hstate *h;
2669         int err;
2670 
2671         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2672         if (!hugepages_kobj)
2673                 return;
2674 
2675         for_each_hstate(h) {
2676                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2677                                          hstate_kobjs, &hstate_attr_group);
2678                 if (err)
2679                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2680         }
2681 }
2682 
2683 #ifdef CONFIG_NUMA
2684 
2685 /*
2686  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2687  * with node devices in node_devices[] using a parallel array.  The array
2688  * index of a node device or _hstate == node id.
2689  * This is here to avoid any static dependency of the node device driver, in
2690  * the base kernel, on the hugetlb module.
2691  */
2692 struct node_hstate {
2693         struct kobject          *hugepages_kobj;
2694         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2695 };
2696 static struct node_hstate node_hstates[MAX_NUMNODES];
2697 
2698 /*
2699  * A subset of global hstate attributes for node devices
2700  */
2701 static struct attribute *per_node_hstate_attrs[] = {
2702         &nr_hugepages_attr.attr,
2703         &free_hugepages_attr.attr,
2704         &surplus_hugepages_attr.attr,
2705         NULL,
2706 };
2707 
2708 static const struct attribute_group per_node_hstate_attr_group = {
2709         .attrs = per_node_hstate_attrs,
2710 };
2711 
2712 /*
2713  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2714  * Returns node id via non-NULL nidp.
2715  */
2716 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2717 {
2718         int nid;
2719 
2720         for (nid = 0; nid < nr_node_ids; nid++) {
2721                 struct node_hstate *nhs = &node_hstates[nid];
2722                 int i;
2723                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2724                         if (nhs->hstate_kobjs[i] == kobj) {
2725                                 if (nidp)
2726                                         *nidp = nid;
2727                                 return &hstates[i];
2728                         }
2729         }
2730 
2731         BUG();
2732         return NULL;
2733 }
2734 
2735 /*
2736  * Unregister hstate attributes from a single node device.
2737  * No-op if no hstate attributes attached.
2738  */
2739 static void hugetlb_unregister_node(struct node *node)
2740 {
2741         struct hstate *h;
2742         struct node_hstate *nhs = &node_hstates[node->dev.id];
2743 
2744         if (!nhs->hugepages_kobj)
2745                 return;         /* no hstate attributes */
2746 
2747         for_each_hstate(h) {
2748                 int idx = hstate_index(h);
2749                 if (nhs->hstate_kobjs[idx]) {
2750                         kobject_put(nhs->hstate_kobjs[idx]);
2751                         nhs->hstate_kobjs[idx] = NULL;
2752                 }
2753         }
2754 
2755         kobject_put(nhs->hugepages_kobj);
2756         nhs->hugepages_kobj = NULL;
2757 }
2758 
2759 
2760 /*
2761  * Register hstate attributes for a single node device.
2762  * No-op if attributes already registered.
2763  */
2764 static void hugetlb_register_node(struct node *node)
2765 {
2766         struct hstate *h;
2767         struct node_hstate *nhs = &node_hstates[node->dev.id];
2768         int err;
2769 
2770         if (nhs->hugepages_kobj)
2771                 return;         /* already allocated */
2772 
2773         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2774                                                         &node->dev.kobj);
2775         if (!nhs->hugepages_kobj)
2776                 return;
2777 
2778         for_each_hstate(h) {
2779                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2780                                                 nhs->hstate_kobjs,
2781                                                 &per_node_hstate_attr_group);
2782                 if (err) {
2783                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2784                                 h->name, node->dev.id);
2785                         hugetlb_unregister_node(node);
2786                         break;
2787                 }
2788         }
2789 }
2790 
2791 /*
2792  * hugetlb init time:  register hstate attributes for all registered node
2793  * devices of nodes that have memory.  All on-line nodes should have
2794  * registered their associated device by this time.
2795  */
2796 static void __init hugetlb_register_all_nodes(void)
2797 {
2798         int nid;
2799 
2800         for_each_node_state(nid, N_MEMORY) {
2801                 struct node *node = node_devices[nid];
2802                 if (node->dev.id == nid)
2803                         hugetlb_register_node(node);
2804         }
2805 
2806         /*
2807          * Let the node device driver know we're here so it can
2808          * [un]register hstate attributes on node hotplug.
2809          */
2810         register_hugetlbfs_with_node(hugetlb_register_node,
2811                                      hugetlb_unregister_node);
2812 }
2813 #else   /* !CONFIG_NUMA */
2814 
2815 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2816 {
2817         BUG();
2818         if (nidp)
2819                 *nidp = -1;
2820         return NULL;
2821 }
2822 
2823 static void hugetlb_register_all_nodes(void) { }
2824 
2825 #endif
2826 
2827 static int __init hugetlb_init(void)
2828 {
2829         int i;
2830 
2831         if (!hugepages_supported())
2832                 return 0;
2833 
2834         if (!size_to_hstate(default_hstate_size)) {
2835                 if (default_hstate_size != 0) {
2836                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2837                                default_hstate_size, HPAGE_SIZE);
2838                 }
2839 
2840                 default_hstate_size = HPAGE_SIZE;
2841                 if (!size_to_hstate(default_hstate_size))
2842                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2843         }
2844         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2845         if (default_hstate_max_huge_pages) {
2846                 if (!default_hstate.max_huge_pages)
2847                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2848         }
2849 
2850         hugetlb_init_hstates();
2851         gather_bootmem_prealloc();
2852         report_hugepages();
2853 
2854         hugetlb_sysfs_init();
2855         hugetlb_register_all_nodes();
2856         hugetlb_cgroup_file_init();
2857 
2858 #ifdef CONFIG_SMP
2859         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2860 #else
2861         num_fault_mutexes = 1;
2862 #endif
2863         hugetlb_fault_mutex_table =
2864                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2865                               GFP_KERNEL);
2866         BUG_ON(!hugetlb_fault_mutex_table);
2867 
2868         for (i = 0; i < num_fault_mutexes; i++)
2869                 mutex_init(&hugetlb_fault_mutex_table[i]);
2870         return 0;
2871 }
2872 subsys_initcall(hugetlb_init);
2873 
2874 /* Should be called on processing a hugepagesz=... option */
2875 void __init hugetlb_bad_size(void)
2876 {
2877         parsed_valid_hugepagesz = false;
2878 }
2879 
2880 void __init hugetlb_add_hstate(unsigned int order)
2881 {
2882         struct hstate *h;
2883         unsigned long i;
2884 
2885         if (size_to_hstate(PAGE_SIZE << order)) {
2886                 pr_warn("hugepagesz= specified twice, ignoring\n");
2887                 return;
2888         }
2889         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2890         BUG_ON(order == 0);
2891         h = &hstates[hugetlb_max_hstate++];
2892         h->order = order;
2893         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2894         h->nr_huge_pages = 0;
2895         h->free_huge_pages = 0;
2896         for (i = 0; i < MAX_NUMNODES; ++i)
2897                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2898         INIT_LIST_HEAD(&h->hugepage_activelist);
2899         h->next_nid_to_alloc = first_memory_node;
2900         h->next_nid_to_free = first_memory_node;
2901         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2902                                         huge_page_size(h)/1024);
2903 
2904         parsed_hstate = h;
2905 }
2906 
2907 static int __init hugetlb_nrpages_setup(char *s)
2908 {
2909         unsigned long *mhp;
2910         static unsigned long *last_mhp;
2911 
2912         if (!parsed_valid_hugepagesz) {
2913                 pr_warn("hugepages = %s preceded by "
2914                         "an unsupported hugepagesz, ignoring\n", s);
2915                 parsed_valid_hugepagesz = true;
2916                 return 1;
2917         }
2918         /*
2919          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2920          * so this hugepages= parameter goes to the "default hstate".
2921          */
2922         else if (!hugetlb_max_hstate)
2923                 mhp = &default_hstate_max_huge_pages;
2924         else
2925                 mhp = &parsed_hstate->max_huge_pages;
2926 
2927         if (mhp == last_mhp) {
2928                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2929                 return 1;
2930         }
2931 
2932         if (sscanf(s, "%lu", mhp) <= 0)
2933                 *mhp = 0;
2934 
2935         /*
2936          * Global state is always initialized later in hugetlb_init.
2937          * But we need to allocate >= MAX_ORDER hstates here early to still
2938          * use the bootmem allocator.
2939          */
2940         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2941                 hugetlb_hstate_alloc_pages(parsed_hstate);
2942 
2943         last_mhp = mhp;
2944 
2945         return 1;
2946 }
2947 __setup("hugepages=", hugetlb_nrpages_setup);
2948 
2949 static int __init hugetlb_default_setup(char *s)
2950 {
2951         default_hstate_size = memparse(s, &s);
2952         return 1;
2953 }
2954 __setup("default_hugepagesz=", hugetlb_default_setup);
2955 
2956 static unsigned int cpuset_mems_nr(unsigned int *array)
2957 {
2958         int node;
2959         unsigned int nr = 0;
2960 
2961         for_each_node_mask(node, cpuset_current_mems_allowed)
2962                 nr += array[node];
2963 
2964         return nr;
2965 }
2966 
2967 #ifdef CONFIG_SYSCTL
2968 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2969                          struct ctl_table *table, int write,
2970                          void __user *buffer, size_t *length, loff_t *ppos)
2971 {
2972         struct hstate *h = &default_hstate;
2973         unsigned long tmp = h->max_huge_pages;
2974         int ret;
2975 
2976         if (!hugepages_supported())
2977                 return -EOPNOTSUPP;
2978 
2979         table->data = &tmp;
2980         table->maxlen = sizeof(unsigned long);
2981         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2982         if (ret)
2983                 goto out;
2984 
2985         if (write)
2986                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2987                                                   NUMA_NO_NODE, tmp, *length);
2988 out:
2989         return ret;
2990 }
2991 
2992 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2993                           void __user *buffer, size_t *length, loff_t *ppos)
2994 {
2995 
2996         return hugetlb_sysctl_handler_common(false, table, write,
2997                                                         buffer, length, ppos);
2998 }
2999 
3000 #ifdef CONFIG_NUMA
3001 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3002                           void __user *buffer, size_t *length, loff_t *ppos)
3003 {
3004         return hugetlb_sysctl_handler_common(true, table, write,
3005                                                         buffer, length, ppos);
3006 }
3007 #endif /* CONFIG_NUMA */
3008 
3009 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3010                         void __user *buffer,
3011                         size_t *length, loff_t *ppos)
3012 {
3013         struct hstate *h = &default_hstate;
3014         unsigned long tmp;
3015         int ret;
3016 
3017         if (!hugepages_supported())
3018                 return -EOPNOTSUPP;
3019 
3020         tmp = h->nr_overcommit_huge_pages;
3021 
3022         if (write && hstate_is_gigantic(h))
3023                 return -EINVAL;
3024 
3025         table->data = &tmp;
3026         table->maxlen = sizeof(unsigned long);
3027         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3028         if (ret)
3029                 goto out;
3030 
3031         if (write) {
3032                 spin_lock(&hugetlb_lock);
3033                 h->nr_overcommit_huge_pages = tmp;
3034                 spin_unlock(&hugetlb_lock);
3035         }
3036 out:
3037         return ret;
3038 }
3039 
3040 #endif /* CONFIG_SYSCTL */
3041 
3042 void hugetlb_report_meminfo(struct seq_file *m)
3043 {
3044         struct hstate *h;
3045         unsigned long total = 0;
3046 
3047         if (!hugepages_supported())
3048                 return;
3049 
3050         for_each_hstate(h) {
3051                 unsigned long count = h->nr_huge_pages;
3052 
3053                 total += (PAGE_SIZE << huge_page_order(h)) * count;
3054 
3055                 if (h == &default_hstate)
3056                         seq_printf(m,
3057                                    "HugePages_Total:   %5lu\n"
3058                                    "HugePages_Free:    %5lu\n"
3059                                    "HugePages_Rsvd:    %5lu\n"
3060                                    "HugePages_Surp:    %5lu\n"
3061                                    "Hugepagesize:   %8lu kB\n",
3062                                    count,
3063                                    h->free_huge_pages,
3064                                    h->resv_huge_pages,
3065                                    h->surplus_huge_pages,
3066                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3067         }
3068 
3069         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3070 }
3071 
3072 int hugetlb_report_node_meminfo(int nid, char *buf)
3073 {
3074         struct hstate *h = &default_hstate;
3075         if (!hugepages_supported())
3076                 return 0;
3077         return sprintf(buf,
3078                 "Node %d HugePages_Total: %5u\n"
3079                 "Node %d HugePages_Free:  %5u\n"
3080                 "Node %d HugePages_Surp:  %5u\n",
3081                 nid, h->nr_huge_pages_node[nid],
3082                 nid, h->free_huge_pages_node[nid],
3083                 nid, h->surplus_huge_pages_node[nid]);
3084 }
3085 
3086 void hugetlb_show_meminfo(void)
3087 {
3088         struct hstate *h;
3089         int nid;
3090 
3091         if (!hugepages_supported())
3092                 return;
3093 
3094         for_each_node_state(nid, N_MEMORY)
3095                 for_each_hstate(h)
3096                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3097                                 nid,
3098                                 h->nr_huge_pages_node[nid],
3099                                 h->free_huge_pages_node[nid],
3100                                 h->surplus_huge_pages_node[nid],
3101                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3102 }
3103 
3104 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3105 {
3106         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3107                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3108 }
3109 
3110 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3111 unsigned long hugetlb_total_pages(void)
3112 {
3113         struct hstate *h;
3114         unsigned long nr_total_pages = 0;
3115 
3116         for_each_hstate(h)
3117                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3118         return nr_total_pages;
3119 }
3120 
3121 static int hugetlb_acct_memory(struct hstate *h, long delta)
3122 {
3123         int ret = -ENOMEM;
3124 
3125         spin_lock(&hugetlb_lock);
3126         /*
3127          * When cpuset is configured, it breaks the strict hugetlb page
3128          * reservation as the accounting is done on a global variable. Such
3129          * reservation is completely rubbish in the presence of cpuset because
3130          * the reservation is not checked against page availability for the
3131          * current cpuset. Application can still potentially OOM'ed by kernel
3132          * with lack of free htlb page in cpuset that the task is in.
3133          * Attempt to enforce strict accounting with cpuset is almost
3134          * impossible (or too ugly) because cpuset is too fluid that
3135          * task or memory node can be dynamically moved between cpusets.
3136          *
3137          * The change of semantics for shared hugetlb mapping with cpuset is
3138          * undesirable. However, in order to preserve some of the semantics,
3139          * we fall back to check against current free page availability as
3140          * a best attempt and hopefully to minimize the impact of changing
3141          * semantics that cpuset has.
3142          */
3143         if (delta > 0) {
3144                 if (gather_surplus_pages(h, delta) < 0)
3145                         goto out;
3146 
3147                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3148                         return_unused_surplus_pages(h, delta);
3149                         goto out;
3150                 }
3151         }
3152 
3153         ret = 0;
3154         if (delta < 0)
3155                 return_unused_surplus_pages(h, (unsigned long) -delta);
3156 
3157 out:
3158         spin_unlock(&hugetlb_lock);
3159         return ret;
3160 }
3161 
3162 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3163 {
3164         struct resv_map *resv = vma_resv_map(vma);
3165 
3166         /*
3167          * This new VMA should share its siblings reservation map if present.
3168          * The VMA will only ever have a valid reservation map pointer where
3169          * it is being copied for another still existing VMA.  As that VMA
3170          * has a reference to the reservation map it cannot disappear until
3171          * after this open call completes.  It is therefore safe to take a
3172          * new reference here without additional locking.
3173          */
3174         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3175                 kref_get(&resv->refs);
3176 }
3177 
3178 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3179 {
3180         struct hstate *h = hstate_vma(vma);
3181         struct resv_map *resv = vma_resv_map(vma);
3182         struct hugepage_subpool *spool = subpool_vma(vma);
3183         unsigned long reserve, start, end;
3184         long gbl_reserve;
3185 
3186         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3187                 return;
3188 
3189         start = vma_hugecache_offset(h, vma, vma->vm_start);
3190         end = vma_hugecache_offset(h, vma, vma->vm_end);
3191 
3192         reserve = (end - start) - region_count(resv, start, end);
3193 
3194         kref_put(&resv->refs, resv_map_release);
3195 
3196         if (reserve) {
3197                 /*
3198                  * Decrement reserve counts.  The global reserve count may be
3199                  * adjusted if the subpool has a minimum size.
3200                  */
3201                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3202                 hugetlb_acct_memory(h, -gbl_reserve);
3203         }
3204 }
3205 
3206 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3207 {
3208         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3209                 return -EINVAL;
3210         return 0;
3211 }
3212 
3213 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3214 {
3215         struct hstate *hstate = hstate_vma(vma);
3216 
3217         return 1UL << huge_page_shift(hstate);
3218 }
3219 
3220 /*
3221  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3222  * handle_mm_fault() to try to instantiate regular-sized pages in the
3223  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3224  * this far.
3225  */
3226 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3227 {
3228         BUG();
3229         return 0;
3230 }
3231 
3232 /*
3233  * When a new function is introduced to vm_operations_struct and added
3234  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3235  * This is because under System V memory model, mappings created via
3236  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3237  * their original vm_ops are overwritten with shm_vm_ops.
3238  */
3239 const struct vm_operations_struct hugetlb_vm_ops = {
3240         .fault = hugetlb_vm_op_fault,
3241         .open = hugetlb_vm_op_open,
3242         .close = hugetlb_vm_op_close,
3243         .split = hugetlb_vm_op_split,
3244         .pagesize = hugetlb_vm_op_pagesize,
3245 };
3246 
3247 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3248                                 int writable)
3249 {
3250         pte_t entry;
3251 
3252         if (writable) {
3253                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3254                                          vma->vm_page_prot)));
3255         } else {
3256                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3257                                            vma->vm_page_prot));
3258         }
3259         entry = pte_mkyoung(entry);
3260         entry = pte_mkhuge(entry);
3261         entry = arch_make_huge_pte(entry, vma, page, writable);
3262 
3263         return entry;
3264 }
3265 
3266 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3267                                    unsigned long address, pte_t *ptep)
3268 {
3269         pte_t entry;
3270 
3271         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3272         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3273                 update_mmu_cache(vma, address, ptep);
3274 }
3275 
3276 bool is_hugetlb_entry_migration(pte_t pte)
3277 {
3278         swp_entry_t swp;
3279 
3280         if (huge_pte_none(pte) || pte_present(pte))
3281                 return false;
3282         swp = pte_to_swp_entry(pte);
3283         if (non_swap_entry(swp) && is_migration_entry(swp))
3284                 return true;
3285         else
3286                 return false;
3287 }
3288 
3289 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3290 {
3291         swp_entry_t swp;
3292 
3293         if (huge_pte_none(pte) || pte_present(pte))
3294                 return 0;
3295         swp = pte_to_swp_entry(pte);
3296         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3297                 return 1;
3298         else
3299                 return 0;
3300 }
3301 
3302 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3303                             struct vm_area_struct *vma)
3304 {
3305         pte_t *src_pte, *dst_pte, entry, dst_entry;
3306         struct page *ptepage;
3307         unsigned long addr;
3308         int cow;
3309         struct hstate *h = hstate_vma(vma);
3310         unsigned long sz = huge_page_size(h);
3311         struct mmu_notifier_range range;
3312         int ret = 0;
3313 
3314         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3315 
3316         if (cow) {
3317                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3318                                         vma->vm_start,
3319                                         vma->vm_end);
3320                 mmu_notifier_invalidate_range_start(&range);
3321         }
3322 
3323         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3324                 spinlock_t *src_ptl, *dst_ptl;
3325                 src_pte = huge_pte_offset(src, addr, sz);
3326                 if (!src_pte)
3327                         continue;
3328                 dst_pte = huge_pte_alloc(dst, addr, sz);
3329                 if (!dst_pte) {
3330                         ret = -ENOMEM;
3331                         break;
3332                 }
3333 
3334                 /*
3335                  * If the pagetables are shared don't copy or take references.
3336                  * dst_pte == src_pte is the common case of src/dest sharing.
3337                  *
3338                  * However, src could have 'unshared' and dst shares with
3339                  * another vma.  If dst_pte !none, this implies sharing.
3340                  * Check here before taking page table lock, and once again
3341                  * after taking the lock below.
3342                  */
3343                 dst_entry = huge_ptep_get(dst_pte);
3344                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3345                         continue;
3346 
3347                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3348                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3349                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3350                 entry = huge_ptep_get(src_pte);
3351                 dst_entry = huge_ptep_get(dst_pte);
3352                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3353                         /*
3354                          * Skip if src entry none.  Also, skip in the
3355                          * unlikely case dst entry !none as this implies
3356                          * sharing with another vma.
3357                          */
3358                         ;
3359                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3360                                     is_hugetlb_entry_hwpoisoned(entry))) {
3361                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3362 
3363                         if (is_write_migration_entry(swp_entry) && cow) {
3364                                 /*
3365                                  * COW mappings require pages in both
3366                                  * parent and child to be set to read.
3367                                  */
3368                                 make_migration_entry_read(&swp_entry);
3369                                 entry = swp_entry_to_pte(swp_entry);
3370                                 set_huge_swap_pte_at(src, addr, src_pte,
3371                                                      entry, sz);
3372                         }
3373                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3374                 } else {
3375                         if (cow) {
3376                                 /*
3377                                  * No need to notify as we are downgrading page
3378                                  * table protection not changing it to point
3379                                  * to a new page.
3380                                  *
3381                                  * See Documentation/vm/mmu_notifier.rst
3382                                  */
3383                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3384                         }
3385                         entry = huge_ptep_get(src_pte);
3386                         ptepage = pte_page(entry);
3387                         get_page(ptepage);
3388                         page_dup_rmap(ptepage, true);
3389                         set_huge_pte_at(dst, addr, dst_pte, entry);
3390                         hugetlb_count_add(pages_per_huge_page(h), dst);
3391                 }
3392                 spin_unlock(src_ptl);
3393                 spin_unlock(dst_ptl);
3394         }
3395 
3396         if (cow)
3397                 mmu_notifier_invalidate_range_end(&range);
3398 
3399         return ret;
3400 }
3401 
3402 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3403                             unsigned long start, unsigned long end,
3404                             struct page *ref_page)
3405 {
3406         struct mm_struct *mm = vma->vm_mm;
3407         unsigned long address;
3408         pte_t *ptep;
3409         pte_t pte;
3410         spinlock_t *ptl;
3411         struct page *page;
3412         struct hstate *h = hstate_vma(vma);
3413         unsigned long sz = huge_page_size(h);
3414         struct mmu_notifier_range range;
3415 
3416         WARN_ON(!is_vm_hugetlb_page(vma));
3417         BUG_ON(start & ~huge_page_mask(h));
3418         BUG_ON(end & ~huge_page_mask(h));
3419 
3420         /*
3421          * This is a hugetlb vma, all the pte entries should point
3422          * to huge page.
3423          */
3424         tlb_change_page_size(tlb, sz);
3425         tlb_start_vma(tlb, vma);
3426 
3427         /*
3428          * If sharing possible, alert mmu notifiers of worst case.
3429          */
3430         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3431                                 end);
3432         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3433         mmu_notifier_invalidate_range_start(&range);
3434         address = start;
3435         for (; address < end; address += sz) {
3436                 ptep = huge_pte_offset(mm, address, sz);
3437                 if (!ptep)
3438                         continue;
3439 
3440                 ptl = huge_pte_lock(h, mm, ptep);
3441                 if (huge_pmd_unshare(mm, &address, ptep)) {
3442                         spin_unlock(ptl);
3443                         /*
3444                          * We just unmapped a page of PMDs by clearing a PUD.
3445                          * The caller's TLB flush range should cover this area.
3446                          */
3447                         continue;
3448                 }
3449 
3450                 pte = huge_ptep_get(ptep);
3451                 if (huge_pte_none(pte)) {
3452                         spin_unlock(ptl);
3453                         continue;
3454                 }
3455 
3456                 /*
3457                  * Migrating hugepage or HWPoisoned hugepage is already
3458                  * unmapped and its refcount is dropped, so just clear pte here.
3459                  */
3460                 if (unlikely(!pte_present(pte))) {
3461                         huge_pte_clear(mm, address, ptep, sz);
3462                         spin_unlock(ptl);
3463                         continue;
3464                 }
3465 
3466                 page = pte_page(pte);
3467                 /*
3468                  * If a reference page is supplied, it is because a specific
3469                  * page is being unmapped, not a range. Ensure the page we
3470                  * are about to unmap is the actual page of interest.
3471                  */
3472                 if (ref_page) {
3473                         if (page != ref_page) {
3474                                 spin_unlock(ptl);
3475                                 continue;
3476                         }
3477                         /*
3478                          * Mark the VMA as having unmapped its page so that
3479                          * future faults in this VMA will fail rather than
3480                          * looking like data was lost
3481                          */
3482                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3483                 }
3484 
3485                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3486                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3487                 if (huge_pte_dirty(pte))
3488                         set_page_dirty(page);
3489 
3490                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3491                 page_remove_rmap(page, true);
3492 
3493                 spin_unlock(ptl);
3494                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3495                 /*
3496                  * Bail out after unmapping reference page if supplied
3497                  */
3498                 if (ref_page)
3499                         break;
3500         }
3501         mmu_notifier_invalidate_range_end(&range);
3502         tlb_end_vma(tlb, vma);
3503 }
3504 
3505 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3506                           struct vm_area_struct *vma, unsigned long start,
3507                           unsigned long end, struct page *ref_page)
3508 {
3509         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3510 
3511         /*
3512          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3513          * test will fail on a vma being torn down, and not grab a page table
3514          * on its way out.  We're lucky that the flag has such an appropriate
3515          * name, and can in fact be safely cleared here. We could clear it
3516          * before the __unmap_hugepage_range above, but all that's necessary
3517          * is to clear it before releasing the i_mmap_rwsem. This works
3518          * because in the context this is called, the VMA is about to be
3519          * destroyed and the i_mmap_rwsem is held.
3520          */
3521         vma->vm_flags &= ~VM_MAYSHARE;
3522 }
3523 
3524 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3525                           unsigned long end, struct page *ref_page)
3526 {
3527         struct mm_struct *mm;
3528         struct mmu_gather tlb;
3529         unsigned long tlb_start = start;
3530         unsigned long tlb_end = end;
3531 
3532         /*
3533          * If shared PMDs were possibly used within this vma range, adjust
3534          * start/end for worst case tlb flushing.
3535          * Note that we can not be sure if PMDs are shared until we try to
3536          * unmap pages.  However, we want to make sure TLB flushing covers
3537          * the largest possible range.
3538          */
3539         adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3540 
3541         mm = vma->vm_mm;
3542 
3543         tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3544         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3545         tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3546 }
3547 
3548 /*
3549  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3550  * mappping it owns the reserve page for. The intention is to unmap the page
3551  * from other VMAs and let the children be SIGKILLed if they are faulting the
3552  * same region.
3553  */
3554 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3555                               struct page *page, unsigned long address)
3556 {
3557         struct hstate *h = hstate_vma(vma);
3558         struct vm_area_struct *iter_vma;
3559         struct address_space *mapping;
3560         pgoff_t pgoff;
3561 
3562         /*
3563          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3564          * from page cache lookup which is in HPAGE_SIZE units.
3565          */
3566         address = address & huge_page_mask(h);
3567         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3568                         vma->vm_pgoff;
3569         mapping = vma->vm_file->f_mapping;
3570 
3571         /*
3572          * Take the mapping lock for the duration of the table walk. As
3573          * this mapping should be shared between all the VMAs,
3574          * __unmap_hugepage_range() is called as the lock is already held
3575          */
3576         i_mmap_lock_write(mapping);
3577         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3578                 /* Do not unmap the current VMA */
3579                 if (iter_vma == vma)
3580                         continue;
3581 
3582                 /*
3583                  * Shared VMAs have their own reserves and do not affect
3584                  * MAP_PRIVATE accounting but it is possible that a shared
3585                  * VMA is using the same page so check and skip such VMAs.
3586                  */
3587                 if (iter_vma->vm_flags & VM_MAYSHARE)
3588                         continue;
3589 
3590                 /*
3591                  * Unmap the page from other VMAs without their own reserves.
3592                  * They get marked to be SIGKILLed if they fault in these
3593                  * areas. This is because a future no-page fault on this VMA
3594                  * could insert a zeroed page instead of the data existing
3595                  * from the time of fork. This would look like data corruption
3596                  */
3597                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3598                         unmap_hugepage_range(iter_vma, address,
3599                                              address + huge_page_size(h), page);
3600         }
3601         i_mmap_unlock_write(mapping);
3602 }
3603 
3604 /*
3605  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3606  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3607  * cannot race with other handlers or page migration.
3608  * Keep the pte_same checks anyway to make transition from the mutex easier.
3609  */
3610 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3611                        unsigned long address, pte_t *ptep,
3612                        struct page *pagecache_page, spinlock_t *ptl)
3613 {
3614         pte_t pte;
3615         struct hstate *h = hstate_vma(vma);
3616         struct page *old_page, *new_page;
3617         int outside_reserve = 0;
3618         vm_fault_t ret = 0;
3619         unsigned long haddr = address & huge_page_mask(h);
3620         struct mmu_notifier_range range;
3621 
3622         pte = huge_ptep_get(ptep);
3623         old_page = pte_page(pte);
3624 
3625 retry_avoidcopy:
3626         /* If no-one else is actually using this page, avoid the copy
3627          * and just make the page writable */
3628         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3629                 page_move_anon_rmap(old_page, vma);
3630                 set_huge_ptep_writable(vma, haddr, ptep);
3631                 return 0;
3632         }
3633 
3634         /*
3635          * If the process that created a MAP_PRIVATE mapping is about to
3636          * perform a COW due to a shared page count, attempt to satisfy
3637          * the allocation without using the existing reserves. The pagecache
3638          * page is used to determine if the reserve at this address was
3639          * consumed or not. If reserves were used, a partial faulted mapping
3640          * at the time of fork() could consume its reserves on COW instead
3641          * of the full address range.
3642          */
3643         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3644                         old_page != pagecache_page)
3645                 outside_reserve = 1;
3646 
3647         get_page(old_page);
3648 
3649         /*
3650          * Drop page table lock as buddy allocator may be called. It will
3651          * be acquired again before returning to the caller, as expected.
3652          */
3653         spin_unlock(ptl);
3654         new_page = alloc_huge_page(vma, haddr, outside_reserve);
3655 
3656         if (IS_ERR(new_page)) {
3657                 /*
3658                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3659                  * it is due to references held by a child and an insufficient
3660                  * huge page pool. To guarantee the original mappers
3661                  * reliability, unmap the page from child processes. The child
3662                  * may get SIGKILLed if it later faults.
3663                  */
3664                 if (outside_reserve) {
3665                         put_page(old_page);
3666                         BUG_ON(huge_pte_none(pte));
3667                         unmap_ref_private(mm, vma, old_page, haddr);
3668                         BUG_ON(huge_pte_none(pte));
3669                         spin_lock(ptl);
3670                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3671                         if (likely(ptep &&
3672                                    pte_same(huge_ptep_get(ptep), pte)))
3673                                 goto retry_avoidcopy;
3674                         /*
3675                          * race occurs while re-acquiring page table
3676                          * lock, and our job is done.
3677                          */
3678                         return 0;
3679                 }
3680 
3681                 ret = vmf_error(PTR_ERR(new_page));
3682                 goto out_release_old;
3683         }
3684 
3685         /*
3686          * When the original hugepage is shared one, it does not have
3687          * anon_vma prepared.
3688          */
3689         if (unlikely(anon_vma_prepare(vma))) {
3690                 ret = VM_FAULT_OOM;
3691                 goto out_release_all;
3692         }
3693 
3694         copy_user_huge_page(new_page, old_page, address, vma,
3695                             pages_per_huge_page(h));
3696         __SetPageUptodate(new_page);
3697 
3698         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3699                                 haddr + huge_page_size(h));
3700         mmu_notifier_invalidate_range_start(&range);
3701 
3702         /*
3703          * Retake the page table lock to check for racing updates
3704          * before the page tables are altered
3705          */
3706         spin_lock(ptl);
3707         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3708         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3709                 ClearPagePrivate(new_page);
3710 
3711                 /* Break COW */
3712                 huge_ptep_clear_flush(vma, haddr, ptep);
3713                 mmu_notifier_invalidate_range(mm, range.start, range.end);
3714                 set_huge_pte_at(mm, haddr, ptep,
3715                                 make_huge_pte(vma, new_page, 1));
3716                 page_remove_rmap(old_page, true);
3717                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3718                 set_page_huge_active(new_page);
3719                 /* Make the old page be freed below */
3720                 new_page = old_page;
3721         }
3722         spin_unlock(ptl);
3723         mmu_notifier_invalidate_range_end(&range);
3724 out_release_all:
3725         restore_reserve_on_error(h, vma, haddr, new_page);
3726         put_page(new_page);
3727 out_release_old:
3728         put_page(old_page);
3729 
3730         spin_lock(ptl); /* Caller expects lock to be held */
3731         return ret;
3732 }
3733 
3734 /* Return the pagecache page at a given address within a VMA */
3735 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3736                         struct vm_area_struct *vma, unsigned long address)
3737 {
3738         struct address_space *mapping;
3739         pgoff_t idx;
3740 
3741         mapping = vma->vm_file->f_mapping;
3742         idx = vma_hugecache_offset(h, vma, address);
3743 
3744         return find_lock_page(mapping, idx);
3745 }
3746 
3747 /*
3748  * Return whether there is a pagecache page to back given address within VMA.
3749  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3750  */
3751 static bool hugetlbfs_pagecache_present(struct hstate *h,
3752                         struct vm_area_struct *vma, unsigned long address)
3753 {
3754         struct address_space *mapping;
3755         pgoff_t idx;
3756         struct page *page;
3757 
3758         mapping = vma->vm_file->f_mapping;
3759         idx = vma_hugecache_offset(h, vma, address);
3760 
3761         page = find_get_page(mapping, idx);
3762         if (page)
3763                 put_page(page);
3764         return page != NULL;
3765 }
3766 
3767 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3768                            pgoff_t idx)
3769 {
3770         struct inode *inode = mapping->host;
3771         struct hstate *h = hstate_inode(inode);
3772         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3773 
3774         if (err)
3775                 return err;
3776         ClearPagePrivate(page);
3777 
3778         /*
3779          * set page dirty so that it will not be removed from cache/file
3780          * by non-hugetlbfs specific code paths.
3781          */
3782         set_page_dirty(page);
3783 
3784         spin_lock(&inode->i_lock);
3785         inode->i_blocks += blocks_per_huge_page(h);
3786         spin_unlock(&inode->i_lock);
3787         return 0;
3788 }
3789 
3790 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3791                         struct vm_area_struct *vma,
3792                         struct address_space *mapping, pgoff_t idx,
3793                         unsigned long address, pte_t *ptep, unsigned int flags)
3794 {
3795         struct hstate *h = hstate_vma(vma);
3796         vm_fault_t ret = VM_FAULT_SIGBUS;
3797         int anon_rmap = 0;
3798         unsigned long size;
3799         struct page *page;
3800         pte_t new_pte;
3801         spinlock_t *ptl;
3802         unsigned long haddr = address & huge_page_mask(h);
3803         bool new_page = false;
3804 
3805         /*
3806          * Currently, we are forced to kill the process in the event the
3807          * original mapper has unmapped pages from the child due to a failed
3808          * COW. Warn that such a situation has occurred as it may not be obvious
3809          */
3810         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3811                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3812                            current->pid);
3813                 return ret;
3814         }
3815 
3816         /*
3817          * Use page lock to guard against racing truncation
3818          * before we get page_table_lock.
3819          */
3820 retry:
3821         page = find_lock_page(mapping, idx);
3822         if (!page) {
3823                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3824                 if (idx >= size)
3825                         goto out;
3826 
3827                 /*
3828                  * Check for page in userfault range
3829                  */
3830                 if (userfaultfd_missing(vma)) {
3831                         u32 hash;
3832                         struct vm_fault vmf = {
3833                                 .vma = vma,
3834                                 .address = haddr,
3835                                 .flags = flags,
3836                                 /*
3837                                  * Hard to debug if it ends up being
3838                                  * used by a callee that assumes
3839                                  * something about the other
3840                                  * uninitialized fields... same as in
3841                                  * memory.c
3842                                  */
3843                         };
3844 
3845                         /*
3846                          * hugetlb_fault_mutex must be dropped before
3847                          * handling userfault.  Reacquire after handling
3848                          * fault to make calling code simpler.
3849                          */
3850                         hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
3851                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3852                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3853                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3854                         goto out;
3855                 }
3856 
3857                 page = alloc_huge_page(vma, haddr, 0);
3858                 if (IS_ERR(page)) {
3859                         ret = vmf_error(PTR_ERR(page));
3860                         goto out;
3861                 }
3862                 clear_huge_page(page, address, pages_per_huge_page(h));
3863                 __SetPageUptodate(page);
3864                 new_page = true;
3865 
3866                 if (vma->vm_flags & VM_MAYSHARE) {
3867                         int err = huge_add_to_page_cache(page, mapping, idx);
3868                         if (err) {
3869                                 put_page(page);
3870                                 if (err == -EEXIST)
3871                                         goto retry;
3872                                 goto out;
3873                         }
3874                 } else {
3875                         lock_page(page);
3876                         if (unlikely(anon_vma_prepare(vma))) {
3877                                 ret = VM_FAULT_OOM;
3878                                 goto backout_unlocked;
3879                         }
3880                         anon_rmap = 1;
3881                 }
3882         } else {
3883                 /*
3884                  * If memory error occurs between mmap() and fault, some process
3885                  * don't have hwpoisoned swap entry for errored virtual address.
3886                  * So we need to block hugepage fault by PG_hwpoison bit check.
3887                  */
3888                 if (unlikely(PageHWPoison(page))) {
3889                         ret = VM_FAULT_HWPOISON |
3890                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3891                         goto backout_unlocked;
3892                 }
3893         }
3894 
3895         /*
3896          * If we are going to COW a private mapping later, we examine the
3897          * pending reservations for this page now. This will ensure that
3898          * any allocations necessary to record that reservation occur outside
3899          * the spinlock.
3900          */
3901         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3902                 if (vma_needs_reservation(h, vma, haddr) < 0) {
3903                         ret = VM_FAULT_OOM;
3904                         goto backout_unlocked;
3905                 }
3906                 /* Just decrements count, does not deallocate */
3907                 vma_end_reservation(h, vma, haddr);
3908         }
3909 
3910         ptl = huge_pte_lock(h, mm, ptep);
3911         size = i_size_read(mapping->host) >> huge_page_shift(h);
3912         if (idx >= size)
3913                 goto backout;
3914 
3915         ret = 0;
3916         if (!huge_pte_none(huge_ptep_get(ptep)))
3917                 goto backout;
3918 
3919         if (anon_rmap) {
3920                 ClearPagePrivate(page);
3921                 hugepage_add_new_anon_rmap(page, vma, haddr);
3922         } else
3923                 page_dup_rmap(page, true);
3924         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3925                                 && (vma->vm_flags & VM_SHARED)));
3926         set_huge_pte_at(mm, haddr, ptep, new_pte);
3927 
3928         hugetlb_count_add(pages_per_huge_page(h), mm);
3929         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3930                 /* Optimization, do the COW without a second fault */
3931                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3932         }
3933 
3934         spin_unlock(ptl);
3935 
3936         /*
3937          * Only make newly allocated pages active.  Existing pages found
3938          * in the pagecache could be !page_huge_active() if they have been
3939          * isolated for migration.
3940          */
3941         if (new_page)
3942                 set_page_huge_active(page);
3943 
3944         unlock_page(page);
3945 out:
3946         return ret;
3947 
3948 backout:
3949         spin_unlock(ptl);
3950 backout_unlocked:
3951         unlock_page(page);
3952         restore_reserve_on_error(h, vma, haddr, page);
3953         put_page(page);
3954         goto out;
3955 }
3956 
3957 #ifdef CONFIG_SMP
3958 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3959                             pgoff_t idx, unsigned long address)
3960 {
3961         unsigned long key[2];
3962         u32 hash;
3963 
3964         key[0] = (unsigned long) mapping;
3965         key[1] = idx;
3966 
3967         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3968 
3969         return hash & (num_fault_mutexes - 1);
3970 }
3971 #else
3972 /*
3973  * For uniprocesor systems we always use a single mutex, so just
3974  * return 0 and avoid the hashing overhead.
3975  */
3976 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3977                             pgoff_t idx, unsigned long address)
3978 {
3979         return 0;
3980 }
3981 #endif
3982 
3983 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3984                         unsigned long address, unsigned int flags)
3985 {
3986         pte_t *ptep, entry;
3987         spinlock_t *ptl;
3988         vm_fault_t ret;
3989         u32 hash;
3990         pgoff_t idx;
3991         struct page *page = NULL;
3992         struct page *pagecache_page = NULL;
3993         struct hstate *h = hstate_vma(vma);
3994         struct address_space *mapping;
3995         int need_wait_lock = 0;
3996         unsigned long haddr = address & huge_page_mask(h);
3997 
3998         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3999         if (ptep) {
4000                 entry = huge_ptep_get(ptep);
4001                 if (unlikely(is_hugetlb_entry_migration(entry))) {
4002                         migration_entry_wait_huge(vma, mm, ptep);
4003                         return 0;
4004                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4005                         return VM_FAULT_HWPOISON_LARGE |
4006                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4007         } else {
4008                 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4009                 if (!ptep)
4010                         return VM_FAULT_OOM;
4011         }
4012 
4013         mapping = vma->vm_file->f_mapping;
4014         idx = vma_hugecache_offset(h, vma, haddr);
4015 
4016         /*
4017          * Serialize hugepage allocation and instantiation, so that we don't
4018          * get spurious allocation failures if two CPUs race to instantiate
4019          * the same page in the page cache.
4020          */
4021         hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
4022         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4023 
4024         entry = huge_ptep_get(ptep);
4025         if (huge_pte_none(entry)) {
4026                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4027                 goto out_mutex;
4028         }
4029 
4030         ret = 0;
4031 
4032         /*
4033          * entry could be a migration/hwpoison entry at this point, so this
4034          * check prevents the kernel from going below assuming that we have
4035          * a active hugepage in pagecache. This goto expects the 2nd page fault,
4036          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4037          * handle it.
4038          */
4039         if (!pte_present(entry))
4040                 goto out_mutex;
4041 
4042         /*
4043          * If we are going to COW the mapping later, we examine the pending
4044          * reservations for this page now. This will ensure that any
4045          * allocations necessary to record that reservation occur outside the
4046          * spinlock. For private mappings, we also lookup the pagecache
4047          * page now as it is used to determine if a reservation has been
4048          * consumed.
4049          */
4050         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4051                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4052                         ret = VM_FAULT_OOM;
4053                         goto out_mutex;
4054                 }
4055                 /* Just decrements count, does not deallocate */
4056                 vma_end_reservation(h, vma, haddr);
4057 
4058                 if (!(vma->vm_flags & VM_MAYSHARE))
4059                         pagecache_page = hugetlbfs_pagecache_page(h,
4060                                                                 vma, haddr);
4061         }
4062 
4063         ptl = huge_pte_lock(h, mm, ptep);
4064 
4065         /* Check for a racing update before calling hugetlb_cow */
4066         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4067                 goto out_ptl;
4068 
4069         /*
4070          * hugetlb_cow() requires page locks of pte_page(entry) and
4071          * pagecache_page, so here we need take the former one
4072          * when page != pagecache_page or !pagecache_page.
4073          */
4074         page = pte_page(entry);
4075         if (page != pagecache_page)
4076                 if (!trylock_page(page)) {
4077                         need_wait_lock = 1;
4078                         goto out_ptl;
4079                 }
4080 
4081         get_page(page);
4082 
4083         if (flags & FAULT_FLAG_WRITE) {
4084                 if (!huge_pte_write(entry)) {
4085                         ret = hugetlb_cow(mm, vma, address, ptep,
4086                                           pagecache_page, ptl);
4087                         goto out_put_page;
4088                 }
4089                 entry = huge_pte_mkdirty(entry);
4090         }
4091         entry = pte_mkyoung(entry);
4092         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4093                                                 flags & FAULT_FLAG_WRITE))
4094                 update_mmu_cache(vma, haddr, ptep);
4095 out_put_page:
4096         if (page != pagecache_page)
4097                 unlock_page(page);
4098         put_page(page);
4099 out_ptl:
4100         spin_unlock(ptl);
4101 
4102         if (pagecache_page) {
4103                 unlock_page(pagecache_page);
4104                 put_page(pagecache_page);
4105         }
4106 out_mutex:
4107         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4108         /*
4109          * Generally it's safe to hold refcount during waiting page lock. But
4110          * here we just wait to defer the next page fault to avoid busy loop and
4111          * the page is not used after unlocked before returning from the current
4112          * page fault. So we are safe from accessing freed page, even if we wait
4113          * here without taking refcount.
4114          */
4115         if (need_wait_lock)
4116                 wait_on_page_locked(page);
4117         return ret;
4118 }
4119 
4120 /*
4121  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4122  * modifications for huge pages.
4123  */
4124 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4125                             pte_t *dst_pte,
4126                             struct vm_area_struct *dst_vma,
4127                             unsigned long dst_addr,
4128                             unsigned long src_addr,
4129                             struct page **pagep)
4130 {
4131         struct address_space *mapping;
4132         pgoff_t idx;
4133         unsigned long size;
4134         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4135         struct hstate *h = hstate_vma(dst_vma);
4136         pte_t _dst_pte;
4137         spinlock_t *ptl;
4138         int ret;
4139         struct page *page;
4140 
4141         if (!*pagep) {
4142                 ret = -ENOMEM;
4143                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4144                 if (IS_ERR(page))
4145                         goto out;
4146 
4147                 ret = copy_huge_page_from_user(page,
4148                                                 (const void __user *) src_addr,
4149                                                 pages_per_huge_page(h), false);
4150 
4151                 /* fallback to copy_from_user outside mmap_sem */
4152                 if (unlikely(ret)) {
4153                         ret = -ENOENT;
4154                         *pagep = page;
4155                         /* don't free the page */
4156                         goto out;
4157                 }
4158         } else {
4159                 page = *pagep;
4160                 *pagep = NULL;
4161         }
4162 
4163         /*
4164          * The memory barrier inside __SetPageUptodate makes sure that
4165          * preceding stores to the page contents become visible before
4166          * the set_pte_at() write.
4167          */
4168         __SetPageUptodate(page);
4169 
4170         mapping = dst_vma->vm_file->f_mapping;
4171         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4172 
4173         /*
4174          * If shared, add to page cache
4175          */
4176         if (vm_shared) {
4177                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4178                 ret = -EFAULT;
4179                 if (idx >= size)
4180                         goto out_release_nounlock;
4181 
4182                 /*
4183                  * Serialization between remove_inode_hugepages() and
4184                  * huge_add_to_page_cache() below happens through the
4185                  * hugetlb_fault_mutex_table that here must be hold by
4186                  * the caller.
4187                  */
4188                 ret = huge_add_to_page_cache(page, mapping, idx);
4189                 if (ret)
4190                         goto out_release_nounlock;
4191         }
4192 
4193         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4194         spin_lock(ptl);
4195 
4196         /*
4197          * Recheck the i_size after holding PT lock to make sure not
4198          * to leave any page mapped (as page_mapped()) beyond the end
4199          * of the i_size (remove_inode_hugepages() is strict about
4200          * enforcing that). If we bail out here, we'll also leave a
4201          * page in the radix tree in the vm_shared case beyond the end
4202          * of the i_size, but remove_inode_hugepages() will take care
4203          * of it as soon as we drop the hugetlb_fault_mutex_table.
4204          */
4205         size = i_size_read(mapping->host) >> huge_page_shift(h);
4206         ret = -EFAULT;
4207         if (idx >= size)
4208                 goto out_release_unlock;
4209 
4210         ret = -EEXIST;
4211         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4212                 goto out_release_unlock;
4213 
4214         if (vm_shared) {
4215                 page_dup_rmap(page, true);
4216         } else {
4217                 ClearPagePrivate(page);
4218                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4219         }
4220 
4221         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4222         if (dst_vma->vm_flags & VM_WRITE)
4223                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4224         _dst_pte = pte_mkyoung(_dst_pte);
4225 
4226         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4227 
4228         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4229                                         dst_vma->vm_flags & VM_WRITE);
4230         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4231 
4232         /* No need to invalidate - it was non-present before */
4233         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4234 
4235         spin_unlock(ptl);
4236         set_page_huge_active(page);
4237         if (vm_shared)
4238                 unlock_page(page);
4239         ret = 0;
4240 out:
4241         return ret;
4242 out_release_unlock:
4243         spin_unlock(ptl);
4244         if (vm_shared)
4245                 unlock_page(page);
4246 out_release_nounlock:
4247         put_page(page);
4248         goto out;
4249 }
4250 
4251 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4252                          struct page **pages, struct vm_area_struct **vmas,
4253                          unsigned long *position, unsigned long *nr_pages,
4254                          long i, unsigned int flags, int *nonblocking)
4255 {
4256         unsigned long pfn_offset;
4257         unsigned long vaddr = *position;
4258         unsigned long remainder = *nr_pages;
4259         struct hstate *h = hstate_vma(vma);
4260         int err = -EFAULT;
4261 
4262         while (vaddr < vma->vm_end && remainder) {
4263                 pte_t *pte;
4264                 spinlock_t *ptl = NULL;
4265                 int absent;
4266                 struct page *page;
4267 
4268                 /*
4269                  * If we have a pending SIGKILL, don't keep faulting pages and
4270                  * potentially allocating memory.
4271                  */
4272                 if (fatal_signal_pending(current)) {
4273                         remainder = 0;
4274                         break;
4275                 }
4276 
4277                 /*
4278                  * Some archs (sparc64, sh*) have multiple pte_ts to
4279                  * each hugepage.  We have to make sure we get the
4280                  * first, for the page indexing below to work.
4281                  *
4282                  * Note that page table lock is not held when pte is null.
4283                  */
4284                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4285                                       huge_page_size(h));
4286                 if (pte)
4287                         ptl = huge_pte_lock(h, mm, pte);
4288                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4289 
4290                 /*
4291                  * When coredumping, it suits get_dump_page if we just return
4292                  * an error where there's an empty slot with no huge pagecache
4293                  * to back it.  This way, we avoid allocating a hugepage, and
4294                  * the sparse dumpfile avoids allocating disk blocks, but its
4295                  * huge holes still show up with zeroes where they need to be.
4296                  */
4297                 if (absent && (flags & FOLL_DUMP) &&
4298                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4299                         if (pte)
4300                                 spin_unlock(ptl);
4301                         remainder = 0;
4302                         break;
4303                 }
4304 
4305                 /*
4306                  * We need call hugetlb_fault for both hugepages under migration
4307                  * (in which case hugetlb_fault waits for the migration,) and
4308                  * hwpoisoned hugepages (in which case we need to prevent the
4309                  * caller from accessing to them.) In order to do this, we use
4310                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4311                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4312                  * both cases, and because we can't follow correct pages
4313                  * directly from any kind of swap entries.
4314                  */
4315                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4316                     ((flags & FOLL_WRITE) &&
4317                       !huge_pte_write(huge_ptep_get(pte)))) {
4318                         vm_fault_t ret;
4319                         unsigned int fault_flags = 0;
4320 
4321                         if (pte)
4322                                 spin_unlock(ptl);
4323                         if (flags & FOLL_WRITE)
4324                                 fault_flags |= FAULT_FLAG_WRITE;
4325                         if (nonblocking)
4326                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4327                         if (flags & FOLL_NOWAIT)
4328                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4329                                         FAULT_FLAG_RETRY_NOWAIT;
4330                         if (flags & FOLL_TRIED) {
4331                                 VM_WARN_ON_ONCE(fault_flags &
4332                                                 FAULT_FLAG_ALLOW_RETRY);
4333                                 fault_flags |= FAULT_FLAG_TRIED;
4334                         }
4335                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4336                         if (ret & VM_FAULT_ERROR) {
4337                                 err = vm_fault_to_errno(ret, flags);
4338                                 remainder = 0;
4339                                 break;
4340                         }
4341                         if (ret & VM_FAULT_RETRY) {
4342                                 if (nonblocking &&
4343                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4344                                         *nonblocking = 0;
4345                                 *nr_pages = 0;
4346                                 /*
4347                                  * VM_FAULT_RETRY must not return an
4348                                  * error, it will return zero
4349                                  * instead.
4350                                  *
4351                                  * No need to update "position" as the
4352                                  * caller will not check it after
4353                                  * *nr_pages is set to 0.
4354                                  */
4355                                 return i;
4356                         }
4357                         continue;
4358                 }
4359 
4360                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4361                 page = pte_page(huge_ptep_get(pte));
4362 
4363                 /*
4364                  * Instead of doing 'try_get_page()' below in the same_page
4365                  * loop, just check the count once here.
4366                  */
4367                 if (unlikely(page_count(page) <= 0)) {
4368                         if (pages) {
4369                                 spin_unlock(ptl);
4370                                 remainder = 0;
4371                                 err = -ENOMEM;
4372                                 break;
4373                         }
4374                 }
4375 same_page:
4376                 if (pages) {
4377                         pages[i] = mem_map_offset(page, pfn_offset);
4378                         get_page(pages[i]);
4379                 }
4380 
4381                 if (vmas)
4382                         vmas[i] = vma;
4383 
4384                 vaddr += PAGE_SIZE;
4385                 ++pfn_offset;
4386                 --remainder;
4387                 ++i;
4388                 if (vaddr < vma->vm_end && remainder &&
4389                                 pfn_offset < pages_per_huge_page(h)) {
4390                         /*
4391                          * We use pfn_offset to avoid touching the pageframes
4392                          * of this compound page.
4393                          */
4394                         goto same_page;
4395                 }
4396                 spin_unlock(ptl);
4397         }
4398         *nr_pages = remainder;
4399         /*
4400          * setting position is actually required only if remainder is
4401          * not zero but it's faster not to add a "if (remainder)"
4402          * branch.
4403          */
4404         *position = vaddr;
4405 
4406         return i ? i : err;
4407 }
4408 
4409 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4410 /*
4411  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4412  * implement this.
4413  */
4414 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4415 #endif
4416 
4417 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4418                 unsigned long address, unsigned long end, pgprot_t newprot)
4419 {
4420         struct mm_struct *mm = vma->vm_mm;
4421         unsigned long start = address;
4422         pte_t *ptep;
4423         pte_t pte;
4424         struct hstate *h = hstate_vma(vma);
4425         unsigned long pages = 0;
4426         bool shared_pmd = false;
4427         struct mmu_notifier_range range;
4428 
4429         /*
4430          * In the case of shared PMDs, the area to flush could be beyond
4431          * start/end.  Set range.start/range.end to cover the maximum possible
4432          * range if PMD sharing is possible.
4433          */
4434         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4435                                 0, vma, mm, start, end);
4436         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4437 
4438         BUG_ON(address >= end);
4439         flush_cache_range(vma, range.start, range.end);
4440 
4441         mmu_notifier_invalidate_range_start(&range);
4442         i_mmap_lock_write(vma->vm_file->f_mapping);
4443         for (; address < end; address += huge_page_size(h)) {
4444                 spinlock_t *ptl;
4445                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4446                 if (!ptep)
4447                         continue;
4448                 ptl = huge_pte_lock(h, mm, ptep);
4449                 if (huge_pmd_unshare(mm, &address, ptep)) {
4450                         pages++;
4451                         spin_unlock(ptl);
4452                         shared_pmd = true;
4453                         continue;
4454                 }
4455                 pte = huge_ptep_get(ptep);
4456                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4457                         spin_unlock(ptl);
4458                         continue;
4459                 }
4460                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4461                         swp_entry_t entry = pte_to_swp_entry(pte);
4462 
4463                         if (is_write_migration_entry(entry)) {
4464                                 pte_t newpte;
4465 
4466                                 make_migration_entry_read(&entry);
4467                                 newpte = swp_entry_to_pte(entry);
4468                                 set_huge_swap_pte_at(mm, address, ptep,
4469                                                      newpte, huge_page_size(h));
4470                                 pages++;
4471                         }
4472                         spin_unlock(ptl);
4473                         continue;
4474                 }
4475                 if (!huge_pte_none(pte)) {
4476                         pte_t old_pte;
4477 
4478                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4479                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4480                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4481                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4482                         pages++;
4483                 }
4484                 spin_unlock(ptl);
4485         }
4486         /*
4487          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4488          * may have cleared our pud entry and done put_page on the page table:
4489          * once we release i_mmap_rwsem, another task can do the final put_page
4490          * and that page table be reused and filled with junk.  If we actually
4491          * did unshare a page of pmds, flush the range corresponding to the pud.
4492          */
4493         if (shared_pmd)
4494                 flush_hugetlb_tlb_range(vma, range.start, range.end);
4495         else
4496                 flush_hugetlb_tlb_range(vma, start, end);
4497         /*
4498          * No need to call mmu_notifier_invalidate_range() we are downgrading
4499          * page table protection not changing it to point to a new page.
4500          *
4501          * See Documentation/vm/mmu_notifier.rst
4502          */
4503         i_mmap_unlock_write(vma->vm_file->f_mapping);
4504         mmu_notifier_invalidate_range_end(&range);
4505 
4506         return pages << h->order;
4507 }
4508 
4509 int hugetlb_reserve_pages(struct inode *inode,
4510                                         long from, long to,
4511                                         struct vm_area_struct *vma,
4512                                         vm_flags_t vm_flags)
4513 {
4514         long ret, chg;
4515         struct hstate *h = hstate_inode(inode);
4516         struct hugepage_subpool *spool = subpool_inode(inode);
4517         struct resv_map *resv_map;
4518         long gbl_reserve;
4519 
4520         /* This should never happen */
4521         if (from > to) {
4522                 VM_WARN(1, "%s called with a negative range\n", __func__);
4523                 return -EINVAL;
4524         }
4525 
4526         /*
4527          * Only apply hugepage reservation if asked. At fault time, an
4528          * attempt will be made for VM_NORESERVE to allocate a page
4529          * without using reserves
4530          */
4531         if (vm_flags & VM_NORESERVE)
4532                 return 0;
4533 
4534         /*
4535          * Shared mappings base their reservation on the number of pages that
4536          * are already allocated on behalf of the file. Private mappings need
4537          * to reserve the full area even if read-only as mprotect() may be
4538          * called to make the mapping read-write. Assume !vma is a shm mapping
4539          */
4540         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4541                 /*
4542                  * resv_map can not be NULL as hugetlb_reserve_pages is only
4543                  * called for inodes for which resv_maps were created (see
4544                  * hugetlbfs_get_inode).
4545                  */
4546                 resv_map = inode_resv_map(inode);
4547 
4548                 chg = region_chg(resv_map, from, to);
4549 
4550         } else {
4551                 resv_map = resv_map_alloc();
4552                 if (!resv_map)
4553                         return -ENOMEM;
4554 
4555                 chg = to - from;
4556 
4557                 set_vma_resv_map(vma, resv_map);
4558                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4559         }
4560 
4561         if (chg < 0) {
4562                 ret = chg;
4563                 goto out_err;
4564         }
4565 
4566         /*
4567          * There must be enough pages in the subpool for the mapping. If
4568          * the subpool has a minimum size, there may be some global
4569          * reservations already in place (gbl_reserve).
4570          */
4571         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4572         if (gbl_reserve < 0) {
4573                 ret = -ENOSPC;
4574                 goto out_err;
4575         }
4576 
4577         /*
4578          * Check enough hugepages are available for the reservation.
4579          * Hand the pages back to the subpool if there are not
4580          */
4581         ret = hugetlb_acct_memory(h, gbl_reserve);
4582         if (ret < 0) {
4583                 /* put back original number of pages, chg */
4584                 (void)hugepage_subpool_put_pages(spool, chg);
4585                 goto out_err;
4586         }
4587 
4588         /*
4589          * Account for the reservations made. Shared mappings record regions
4590          * that have reservations as they are shared by multiple VMAs.
4591          * When the last VMA disappears, the region map says how much
4592          * the reservation was and the page cache tells how much of
4593          * the reservation was consumed. Private mappings are per-VMA and
4594          * only the consumed reservations are tracked. When the VMA
4595          * disappears, the original reservation is the VMA size and the
4596          * consumed reservations are stored in the map. Hence, nothing
4597          * else has to be done for private mappings here
4598          */
4599         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4600                 long add = region_add(resv_map, from, to);
4601 
4602                 if (unlikely(chg > add)) {
4603                         /*
4604                          * pages in this range were added to the reserve
4605                          * map between region_chg and region_add.  This
4606                          * indicates a race with alloc_huge_page.  Adjust
4607                          * the subpool and reserve counts modified above
4608                          * based on the difference.
4609                          */
4610                         long rsv_adjust;
4611 
4612                         rsv_adjust = hugepage_subpool_put_pages(spool,
4613                                                                 chg - add);
4614                         hugetlb_acct_memory(h, -rsv_adjust);
4615                 }
4616         }
4617         return 0;
4618 out_err:
4619         if (!vma || vma->vm_flags & VM_MAYSHARE)
4620                 /* Don't call region_abort if region_chg failed */
4621                 if (chg >= 0)
4622                         region_abort(resv_map, from, to);
4623         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4624                 kref_put(&resv_map->refs, resv_map_release);
4625         return ret;
4626 }
4627 
4628 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4629                                                                 long freed)
4630 {
4631         struct hstate *h = hstate_inode(inode);
4632         struct resv_map *resv_map = inode_resv_map(inode);
4633         long chg = 0;
4634         struct hugepage_subpool *spool = subpool_inode(inode);
4635         long gbl_reserve;
4636 
4637         /*
4638          * Since this routine can be called in the evict inode path for all
4639          * hugetlbfs inodes, resv_map could be NULL.
4640          */
4641         if (resv_map) {
4642                 chg = region_del(resv_map, start, end);
4643                 /*
4644                  * region_del() can fail in the rare case where a region
4645                  * must be split and another region descriptor can not be
4646                  * allocated.  If end == LONG_MAX, it will not fail.
4647                  */
4648                 if (chg < 0)
4649                         return chg;
4650         }
4651 
4652         spin_lock(&inode->i_lock);
4653         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4654         spin_unlock(&inode->i_lock);
4655 
4656         /*
4657          * If the subpool has a minimum size, the number of global
4658          * reservations to be released may be adjusted.
4659          */
4660         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4661         hugetlb_acct_memory(h, -gbl_reserve);
4662 
4663         return 0;
4664 }
4665 
4666 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4667 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4668                                 struct vm_area_struct *vma,
4669                                 unsigned long addr, pgoff_t idx)
4670 {
4671         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4672                                 svma->vm_start;
4673         unsigned long sbase = saddr & PUD_MASK;
4674         unsigned long s_end = sbase + PUD_SIZE;
4675 
4676         /* Allow segments to share if only one is marked locked */
4677         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4678         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4679 
4680         /*
4681          * match the virtual addresses, permission and the alignment of the
4682          * page table page.
4683          */
4684         if (pmd_index(addr) != pmd_index(saddr) ||
4685             vm_flags != svm_flags ||
4686             sbase < svma->vm_start || svma->vm_end < s_end)
4687                 return 0;
4688 
4689         return saddr;
4690 }
4691 
4692 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4693 {
4694         unsigned long base = addr & PUD_MASK;
4695         unsigned long end = base + PUD_SIZE;
4696 
4697         /*
4698          * check on proper vm_flags and page table alignment
4699          */
4700         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4701                 return true;
4702         return false;
4703 }
4704 
4705 /*
4706  * Determine if start,end range within vma could be mapped by shared pmd.
4707  * If yes, adjust start and end to cover range associated with possible
4708  * shared pmd mappings.
4709  */
4710 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4711                                 unsigned long *start, unsigned long *end)
4712 {
4713         unsigned long check_addr = *start;
4714 
4715         if (!(vma->vm_flags & VM_MAYSHARE))
4716                 return;
4717 
4718         for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4719                 unsigned long a_start = check_addr & PUD_MASK;
4720                 unsigned long a_end = a_start + PUD_SIZE;
4721 
4722                 /*
4723                  * If sharing is possible, adjust start/end if necessary.
4724                  */
4725                 if (range_in_vma(vma, a_start, a_end)) {
4726                         if (a_start < *start)
4727                                 *start = a_start;
4728                         if (a_end > *end)
4729                                 *end = a_end;
4730                 }
4731         }
4732 }
4733 
4734 /*
4735  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4736  * and returns the corresponding pte. While this is not necessary for the
4737  * !shared pmd case because we can allocate the pmd later as well, it makes the
4738  * code much cleaner. pmd allocation is essential for the shared case because
4739  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4740  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4741  * bad pmd for sharing.
4742  */
4743 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4744 {
4745         struct vm_area_struct *vma = find_vma(mm, addr);
4746         struct address_space *mapping = vma->vm_file->f_mapping;
4747         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4748                         vma->vm_pgoff;
4749         struct vm_area_struct *svma;
4750         unsigned long saddr;
4751         pte_t *spte = NULL;
4752         pte_t *pte;
4753         spinlock_t *ptl;
4754 
4755         if (!vma_shareable(vma, addr))
4756                 return (pte_t *)pmd_alloc(mm, pud, addr);
4757 
4758         i_mmap_lock_write(mapping);
4759         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4760                 if (svma == vma)
4761                         continue;
4762 
4763                 saddr = page_table_shareable(svma, vma, addr, idx);
4764                 if (saddr) {
4765                         spte = huge_pte_offset(svma->vm_mm, saddr,
4766                                                vma_mmu_pagesize(svma));
4767                         if (spte) {
4768                                 get_page(virt_to_page(spte));
4769                                 break;
4770                         }
4771                 }
4772         }
4773 
4774         if (!spte)
4775                 goto out;
4776 
4777         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4778         if (pud_none(*pud)) {
4779                 pud_populate(mm, pud,
4780                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4781                 mm_inc_nr_pmds(mm);
4782         } else {
4783                 put_page(virt_to_page(spte));
4784         }
4785         spin_unlock(ptl);
4786 out:
4787         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4788         i_mmap_unlock_write(mapping);
4789         return pte;
4790 }
4791 
4792 /*
4793  * unmap huge page backed by shared pte.
4794  *
4795  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4796  * indicated by page_count > 1, unmap is achieved by clearing pud and
4797  * decrementing the ref count. If count == 1, the pte page is not shared.
4798  *
4799  * called with page table lock held.
4800  *
4801  * returns: 1 successfully unmapped a shared pte page
4802  *          0 the underlying pte page is not shared, or it is the last user
4803  */
4804 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4805 {
4806         pgd_t *pgd = pgd_offset(mm, *addr);
4807         p4d_t *p4d = p4d_offset(pgd, *addr);
4808         pud_t *pud = pud_offset(p4d, *addr);
4809 
4810         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4811         if (page_count(virt_to_page(ptep)) == 1)
4812                 return 0;
4813 
4814         pud_clear(pud);
4815         put_page(virt_to_page(ptep));
4816         mm_dec_nr_pmds(mm);
4817         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4818         return 1;
4819 }
4820 #define want_pmd_share()        (1)
4821 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4822 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4823 {
4824         return NULL;
4825 }
4826 
4827 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4828 {
4829         return 0;
4830 }
4831 
4832 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4833                                 unsigned long *start, unsigned long *end)
4834 {
4835 }
4836 #define want_pmd_share()        (0)
4837 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4838 
4839 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4840 pte_t *huge_pte_alloc(struct mm_struct *mm,
4841                         unsigned long addr, unsigned long sz)
4842 {
4843         pgd_t *pgd;
4844         p4d_t *p4d;
4845         pud_t *pud;
4846         pte_t *pte = NULL;
4847 
4848         pgd = pgd_offset(mm, addr);
4849         p4d = p4d_alloc(mm, pgd, addr);
4850         if (!p4d)
4851                 return NULL;
4852         pud = pud_alloc(mm, p4d, addr);
4853         if (pud) {
4854                 if (sz == PUD_SIZE) {
4855                         pte = (pte_t *)pud;
4856                 } else {
4857                         BUG_ON(sz != PMD_SIZE);
4858                         if (want_pmd_share() && pud_none(*pud))
4859                                 pte = huge_pmd_share(mm, addr, pud);
4860                         else
4861                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4862                 }
4863         }
4864         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4865 
4866         return pte;
4867 }
4868 
4869 /*
4870  * huge_pte_offset() - Walk the page table to resolve the hugepage
4871  * entry at address @addr
4872  *
4873  * Return: Pointer to page table or swap entry (PUD or PMD) for
4874  * address @addr, or NULL if a p*d_none() entry is encountered and the
4875  * size @sz doesn't match the hugepage size at this level of the page
4876  * table.
4877  */
4878 pte_t *huge_pte_offset(struct mm_struct *mm,
4879                        unsigned long addr, unsigned long sz)
4880 {
4881         pgd_t *pgd;
4882         p4d_t *p4d;
4883         pud_t *pud;
4884         pmd_t *pmd;
4885 
4886         pgd = pgd_offset(mm, addr);
4887         if (!pgd_present(*pgd))
4888                 return NULL;
4889         p4d = p4d_offset(pgd, addr);
4890         if (!p4d_present(*p4d))
4891                 return NULL;
4892 
4893         pud = pud_offset(p4d, addr);
4894         if (sz != PUD_SIZE && pud_none(*pud))
4895                 return NULL;
4896         /* hugepage or swap? */
4897         if (pud_huge(*pud) || !pud_present(*pud))
4898                 return (pte_t *)pud;
4899 
4900         pmd = pmd_offset(pud, addr);
4901         if (sz != PMD_SIZE && pmd_none(*pmd))
4902                 return NULL;
4903         /* hugepage or swap? */
4904         if (pmd_huge(*pmd) || !pmd_present(*pmd))
4905                 return (pte_t *)pmd;
4906 
4907         return NULL;
4908 }
4909 
4910 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4911 
4912 /*
4913  * These functions are overwritable if your architecture needs its own
4914  * behavior.
4915  */
4916 struct page * __weak
4917 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4918                               int write)
4919 {
4920         return ERR_PTR(-EINVAL);
4921 }
4922 
4923 struct page * __weak
4924 follow_huge_pd(struct vm_area_struct *vma,
4925                unsigned long address, hugepd_t hpd, int flags, int pdshift)
4926 {
4927         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4928         return NULL;
4929 }
4930 
4931 struct page * __weak
4932 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4933                 pmd_t *pmd, int flags)
4934 {
4935         struct page *page = NULL;
4936         spinlock_t *ptl;
4937         pte_t pte;
4938 retry:
4939         ptl = pmd_lockptr(mm, pmd);
4940         spin_lock(ptl);
4941         /*
4942          * make sure that the address range covered by this pmd is not
4943          * unmapped from other threads.
4944          */
4945         if (!pmd_huge(*pmd))
4946                 goto out;
4947         pte = huge_ptep_get((pte_t *)pmd);
4948         if (pte_present(pte)) {
4949                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4950                 if (flags & FOLL_GET)
4951                         get_page(page);
4952         } else {
4953                 if (is_hugetlb_entry_migration(pte)) {
4954                         spin_unlock(ptl);
4955                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4956                         goto retry;
4957                 }
4958                 /*
4959                  * hwpoisoned entry is treated as no_page_table in
4960                  * follow_page_mask().
4961                  */
4962         }
4963 out:
4964         spin_unlock(ptl);
4965         return page;
4966 }
4967 
4968 struct page * __weak
4969 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4970                 pud_t *pud, int flags)
4971 {
4972         if (flags & FOLL_GET)
4973                 return NULL;
4974 
4975         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4976 }
4977 
4978 struct page * __weak
4979 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4980 {
4981         if (flags & FOLL_GET)
4982                 return NULL;
4983 
4984         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4985 }
4986 
4987 bool isolate_huge_page(struct page *page, struct list_head *list)
4988 {
4989         bool ret = true;
4990 
4991         VM_BUG_ON_PAGE(!PageHead(page), page);
4992         spin_lock(&hugetlb_lock);
4993         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4994                 ret = false;
4995                 goto unlock;
4996         }
4997         clear_page_huge_active(page);
4998         list_move_tail(&page->lru, list);
4999 unlock:
5000         spin_unlock(&hugetlb_lock);
5001         return ret;
5002 }
5003 
5004 void putback_active_hugepage(struct page *page)
5005 {
5006         VM_BUG_ON_PAGE(!PageHead(page), page);
5007         spin_lock(&hugetlb_lock);
5008         set_page_huge_active(page);
5009         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5010         spin_unlock(&hugetlb_lock);
5011         put_page(page);
5012 }
5013 
5014 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5015 {
5016         struct hstate *h = page_hstate(oldpage);
5017 
5018         hugetlb_cgroup_migrate(oldpage, newpage);
5019         set_page_owner_migrate_reason(newpage, reason);
5020 
5021         /*
5022          * transfer temporary state of the new huge page. This is
5023          * reverse to other transitions because the newpage is going to
5024          * be final while the old one will be freed so it takes over
5025          * the temporary status.
5026          *
5027          * Also note that we have to transfer the per-node surplus state
5028          * here as well otherwise the global surplus count will not match
5029          * the per-node's.
5030          */
5031         if (PageHugeTemporary(newpage)) {
5032                 int old_nid = page_to_nid(oldpage);
5033                 int new_nid = page_to_nid(newpage);
5034 
5035                 SetPageHugeTemporary(oldpage);
5036                 ClearPageHugeTemporary(newpage);
5037 
5038                 spin_lock(&hugetlb_lock);
5039                 if (h->surplus_huge_pages_node[old_nid]) {
5040                         h->surplus_huge_pages_node[old_nid]--;
5041                         h->surplus_huge_pages_node[new_nid]++;
5042                 }
5043                 spin_unlock(&hugetlb_lock);
5044         }
5045 }
5046 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp